From Policy to Practice: Built-In CIS Benchmarks on Azure - Flexible, Hybrid-Ready
Security is more important than ever. The industry-standard for secure machine configuration is the Center for Internet Security (CIS) Benchmarks. These benchmarks provide consensus-based prescriptive guidance to help organizations harden diverse systems, reduce risk, and streamline compliance with major regulatory frameworks and industry standards like NIST, HIPAA, and PCI DSS. In our previous post, we outlined our plans to improve the Linux server compliance and hardening experience on Azure and shared a vision for integrating CIS Benchmarks. Today, that vision has turned into reality. We're now announcing the next phase of this work: Center for Internet Security (CIS) Benchmarks are now available on Azure for all Azure endorsed distros, at no additional cost to Azure and Azure Arc customers. With today's announcement, you get access to the CIS Benchmarks on Azure with full parity to what’s published by the Center for Internet Security (CIS). You can adjust parameters or define exceptions, tailoring security to your needs and applying consistent controls across cloud, hybrid, and on-premises environments - without having to implement every control manually. Thanks to this flexible architecture, you can truly manage compliance as code. How we achieve parity To ensure accuracy and trust, we rely on and ingest CIS machine-readable Benchmark content (OVAL/XCCDF files) as the source of truth. This guarantees that the controls and rules you apply in Azure match the official CIS specifications, reducing drift and ensuring compliance confidence. What’s new under the hood At the core of this update is kompli - a lightweight, open-source module developed by the Azure Core Linux team. It evaluates Linux systems directly against industry-standard benchmarks like CIS, supporting both audit and, in the future, auto-remediation. This enables accurate, scalable compliance checks across large Linux fleets. Here you can read more about kompli. Dynamic rule evaluation The new compliance engine supports simple fact-checking operations, evaluation of logic operations on them (e.g., anyOf, allOf) and Lua based scripting, which allows to express complex checks required by the CIS Critical Security Controls - all evaluated natively without external scripts. Scalable architecture for large fleets When the assignment is created, the Azure control plane instructs the machine to pull the latest Policy package via the Machine Configuration agent. kompli is integrated as a light-weight library to the package and called by Machine Configuration agent for evaluation – which happens every 15-30minutes. This ensures near real-time compliance state without overwhelming resources and enables consistent evaluation across thousands of VMs and Azure Arc-enabled servers. Future-ready for remediation and enforcement While the Public Preview starts with audit-only mode, the roadmap includes per-rule remediation and enforcement using technologies like eBPF for kernel-level controls. This will allow proactive prevention of configuration drift and runtime hardening at scale. Please reach out if you interested in auto-remediation or enforcement. Extensibility beyond CIS Benchmarks The architecture was designed to support other security and compliance standards as well and isn’t limited to CIS Benchmarks. The compliance engine is modular, and we plan to extend the platform with STIG and other relevant industry benchmarks. This positions Azure as a platform for a place where you can manage your compliance from a single control-plane without duplicating efforts elsewhere. Collaboration with the CIS This milestone reflects a close collaboration between Microsoft and the CIS to bring industry-standard security guidance into Azure as a built-in capability. Our shared goal is to make cloud-native compliance practical and consistent, while giving customers the flexibility to meet their unique requirements. We are committed to continuously supporting new Benchmark releases, expanding coverage with new distributions and easing adoption through built-in workflows, such as moving from your current Benchmark version to a new version while preserving your custom configurations. Certification and trust We can proudly announce that kompli has met all the requirements and is officially certified by the CIS for Benchmark assessment, so you can trust compliance results as authoritative. Minor benchmark updates will be applied automatically, while major version will be released separately. We will include workflows to help migrate customizations seamlessly across versions. Key Highlights Built-in CIS Benchmarks for Azure Endorsed Linux distributions Full parity with official CIS Benchmarks content and certified by the CIS for Benchmark Assessment Flexible configuration: adjust parameters, define exceptions, tune severity Hybrid support: enforce the same baseline across Azure, on-prem, and multi-cloud with Azure Arc Reporting format in CIS tooling style Supported use cases Certified CIS Benchmarks for all Azure Endorsed Distros - Audit only (L1/L2 server profiles) Hybrid / On-premises and other cloud machines with Azure Arc for the supported distros Compliance as Code (example via Github -> Azure OIDC auth and API integration) Compatible with GuestConfig workbook What’s next? Our next mission is to bring the previously announced auto-remediation capability into this experience, expand the distribution coverage and elevate our workflows even further. We’re focused on empowering you to resolve issues while honoring the unique operational complexity of your environments. Stay tuned! Get Started Documentation link for this capability Enable CIS Benchmarks in Machine Configuration and select the “Official Center for Internet Security (CIS) Benchmarks for Linux Workloads” then select the distributions for your assignment, and customize as needed. In case if you want any additional distribution supported or have any feedback for kompli – please open an Azure support case or a Github issue here Relevant Ignite 2025 session: Hybrid workload compliance from policy to practice on Azure Connect with us at Ignite Meet the Linux team and stop by the Linux on Azure booth to see these innovations in action: Session Type Session Code Session Name Date/Time (PST) Theatre THR 712 Hybrid workload compliance from policy to practice on Azure Tue, Nov 18/ 3:15 PM – 3:45 PM Breakout BRK 143 Optimizing performance, deployments, and security for Linux on Azure Thu, Nov 20/ 1:00 PM – 1:45 PM Breakout BRK 144 Build, modernize, and secure AKS workloads with Azure Linux Wed, Nov 19/ 1:30 PM – 2:15 PM Breakout BRK 104 From VMs and containers to AI apps with Azure Red Hat OpenShift Thu, Nov 20/ 8:30 AM – 9:15 AM Theatre THR 701 From Container to Node: Building Minimal-CVE Solutions with Azure Linux Wed, Nov 19/ 3:30 PM – 4:00 PM Lab Lab 505 Fast track your Linux and PostgreSQL migration with Azure Migrate Tue, Nov 18/ 4:30 PM – 5:45 PM PST Wed, Nov 19/ 3:45 PM – 5:00 PM PST Thu, Nov 20/ 9:00 AM – 10:15 AM PSTAnnouncing the Open-Source Release of ML Video Codec (MLVC)
Video codecs compress video for transmission or storage, reducing bandwidth and storage requirements. MLVC is a modern machine-learning-based codec that uses substantially less bandwidth than conventional codecs, improving streaming and video-call quality—especially on constrained or unreliable networks—while lowering delivery and storage costs. MLVC is the product iteration of DCVC (Deep Contextual Video Compression) family of NVC (Neural Video Codec), open sourced by Microsoft Research since 2021, with improved compression efficiency, real-time performance on commodity Neural Processing Units (NPUs), and cross-platform support. We recently published this work in the paper MLVC: Multi-platform Learned Video Codec for Real-World Deployment. We are releasing the source code because we believe the next generation of video coding will be built openly, and we want the broader community — researchers, video codec engineers, platform vendors, product teams, as well as general developer community — to build it with us. Why MLVC Traditional video codecs (e.g., H.264/AVC, H.265/HEVC) have served the industry for a long time, but each generation requires enormous engineering effort for incremental gains and needs dedicated hardware which takes years to become commonly available. MLVC replaces conventional primitives — motion estimation, transforms, entropy modeling — with end-to-end learned neural compression, trained directly against rate-distortion objectives, and run on general-purpose NPU devices. The table below compares MLVC to popular video codecs, showing its lower bitrate and resulting savings in bandwidth and storage. Resolution vs H.264 vs H.265 360p 87.8% 75.5% 540p 82.7% 65.4% For example, for 360p video at 30 fps, where H.264 requires 1 Mbps, MLVC requires roughly 122 kbps for equivalent quality — about one-eight the bitrate under real-time conditions. The inference compute was kept approximately equal for the 360p and 540p resolutions. These results are based on a P.910 subjective test and are based on the Video Conferencing Dataset (VCD) dataset that we developed and also recently released as an open-source project. The following video demo illustrates the extent of quality enhancement achieved by MLVC relative to H.265/HEVC at the same bitrate of 200kbps (please watch by opening in a new window for better demonstration) \ Beyond video compression efficiency, MLVC also offers: NPU-first design. MLVC is built to run almost entirely on the AI accelerators already shipping in modern devices — Apple Neural Engine, Qualcomm and Intel NPUs — at no more than 50% NPU utilization, leaving NPU headroom for the rest of the system. Real-time execution at the targeted operating points. Demonstrated 540p at 30 fps on Apple, Intel, and Qualcomm hardware. A scaling-law trajectory. Empirically, MLVC's coding efficiency improves with increased model capacity and additional training compute. Content-adaptive behavior out of the box, without hand-tuned Rate Distortion Optimization heuristics. Already running in Microsoft Teams MLVC is more than just a research concept for Microsoft. We are currently rolling it out in Microsoft Teams, where it is being validated on real peer-to-peer video calls with active telemetry and A/B testing. The integration runs alongside fallback to conventional video codecs for hardware or reliability constraints — the kind of mixed-environment deployment that real products need. The scaling and reliability insights from this rollout are shaping the codec and its roadmap. We welcome your contributions We can not cover every use case, every device class, or every content domain by ourselves. That is why we are open sourcing MLVC. If you work on: Streaming, Video On Demand (VOD), or live broadcast Real-time communication and conferencing Cloud gaming or remote rendering Surveillance, drones, or robotics Mobile capture, Augmented Reality (AR) / Virtual Reality (VR), or volumetric video Codec hardware, NPUs, or inference runtimes We would love your help. Contributions of all kinds are welcome, including model improvements, training recipes, new platform ports and conversion targets, runtime backends, domain-specific fine-tunes, evaluation tooling, bug reports, and feedback on what is missing for your scenario. We particularly welcome platform ports that expand NPU coverage and efficiency improvements that push the rate-distortion frontier. What's in the release The MLVC repository is available at https://github.com/microsoft/mlvc and shared under the MIT License. It includes: MLVC model source code of the full network architecture. Trained model weights ready to run. Training scripts used to produce shipping models. Training data collection documentation to help reproduce and improve MLVC. Platform conversion scripts to target different NPUs and runtimes. Issues and pull requests will be open from day one. A follow-up release will add a C++ codec library, simplifying integration into real-world applications. Where we're heading Our long-term goal with MLVC is to create an open, learned video codec that meets or exceeds the coding efficiency of the best conventional codecs across the full range of video content, runs efficiently on the AI hardware already shipping in client and cloud devices, scales with compute the way modern ML systems do, and evolves in the open at the pace of the ML community rather than the pace of standardization cycles. In the near term that means stabilizing 540p real-time performance, expanding hardware coverage, and improving loss resilience. In the medium term: higher resolution, e.g., 1080p, and broader streaming scenarios. In the long term: an open video codec ecosystem that meaningfully replaces legacy stacks where it makes sense to. We can't create the future of MLVC alone. We're glad you're here, and we are looking forward to building the next-generation video codec with you. Who are we MLVC is brought to you by the following awesome folks working on the project at Microsoft: Ross Cutler, Ando Saabas, Tanel Pärnamaa, Ardi Loot, Haiyan Xie, Lauri Ehrenpreis, Andrei Znobishchev, Martin Lumiste, Evgenii Indenbom, Yan Lu, Bin Li, Jiahao Li, Naba Kumar, Babak Naderi, Juhee Cho, Badal Yadav, Jinxin Zhou, Tianyu Ding, Patrick Gregory. — The MLVC Team, MicrosoftClosing the AI Agent Governance Gap with Microsoft Foundry
Developers are shipping agents faster than security teams can catalog them. As organizations move beyond pilots and begin operating dozens or hundreds of agents, one question keeps coming up: how do we actually get visibility into our AI agents across our environment? In this article, we'll walk through how Azure services can help establish visibility, guardrails, and cost accountability across your AI estate. Governance for AI happens across four layers: Resources – who can create new AI resources Builders – who can develop and publish agents in a certain scope Behavior – how agent outputs are evaluated, monitored, and governed Dependencies – what models, tools, APIs, and MCP servers agents can interact with Most organizations already have governance controls for identities, networking, and compliance. The challenge isn't creating new controls. It's connecting existing controls into an operating model that works for AI agents. Below we walk through each area and go a bit deeper on how to close the gap. Setting up boundaries with Azure Policy First, let's start in the Azure portal with Azure Policy. Azure Policy lets you set guardrails on what can be deployed in your environment and flags or blocks anything that doesn't comply. For AI workloads, the built-in definitions range from limiting models that people in your organization can deploy to locking down the network through enabling private endpoints. Some policies you get started with: Foundry model deployments should only use approved models: Restricts deployments to models or publishers your organization has explicitly approved Foundry model deployments should meet eligibility requirements (preview): Applies rules based on model attributes like preview vs. GA status and distribution source Azure AI Services resources should have key access disabled: Makes Microsoft Entra ID the only entry point The full list of policies related to Azure AI Services are available here: List of built-in policy definitions - Azure Policy | Microsoft Learn Why this comes first: Policy checks resources before they are created, so it proactively keeps your environment aligned with your standards. Implementing role-based access control (RBAC) Once these boundaries are in place, the next step is RBAC. Setting RBAC up early ensures that people and identities building agents have the right scope for what they actually need to do. Foundry roles only apply when you authenticate using Microsoft Entra ID. If you're using key-based authentication instead, the key grants full access with without role restrictions. API keys are convenient for quick development usage but when moving towards production, Microsoft Entra ID is the preferred method. Roles can be assigned at three scopes: the Foundry resource, a Foundry project, or an individual agent itself. Below is an example of how different personas within organization can map to a certain scope for creating and building agents with Foundry. Here's how each role in the diagram compares, from least to most privileged: Role Privilege Level What it does in Microsoft Foundry Foundry Agent Consumer Least Interact with agent endpoints in a project. This is your least-privilege role for people who only need to use agents. Foundry User Low Grants reader access to the Foundry project, the Foundry resource, and data actions for your Foundry project. Least-privilege access role for developers building and testing agents. Foundry Project Manager Medium This role lets you perform management actions on Foundry projects, build and develop with projects, and conditionally assign the Foundry User role to other user principals. Foundry Account Owner Higher Grants full access to manage Foundry projects and resources, and lets you assign the Foundry User role to other user principals. Foundry Owner Highest Grants full access to manage Foundry projects and resources to build and develop with projects. This role can also assign the Foundry User, ACR, and monitoring roles to users in the environment. Source: Role-based access control for Microsoft Foundry - Microsoft Foundry | Microsoft Learn For the agent resources themselves, assign managed identities rather than API keys since it lowers the risk of having compromised credentials. Note if you're scripting RBAC permissions: these roles were recently renamed from Azure AI User, Azure AI Owner, Azure AI Account Owner, and Azure AI Project Manager. The role IDs and permissions didn't change, so use the role definition GUID in your code to avoid issues while the rename rolls out. Observability in Microsoft Foundry Governance requires more than access control. Organizations also need evidence of how agents are being used. Observability provides the audit trail needed to investigate incidents, understand usage patterns, and track costs. The Foundry Control Plane brings these observability and governance tools together in one place, alongside services like Azure Monitor, Microsoft Entra, Microsoft Purview, and Azure Policy. In the Foundry portal, tracing is a good starting point. Once you connect an Application Insights resource to your project, Foundry turns on tracing automatically so every run, including the ones you test in the playground, is logged. After it is completed, you can search by Response ID or Trace ID to see the conversation history, token usage, run steps, tool calls, and inputs and outputs between the user and the agent. For more granular queries, you can write KQL to dive into individual agent runs or use the prebuilt Grafana dashboards in Azure Monitor. With client-side tracing, you can also export traces to observability tools you may already use, such as Datadog or Jaeger. Note: Permissions required for viewing this telemetry requires the Log Analytics Reader role on the connected Application Insights resource, and Privileged Monitoring Data Reader on top of that if the underlying Log Analytics tables are protected. Alongside all of these monitoring features, every agent comes with content safety guardrails and evaluations let you test and optimize performance before and after you publish your agents. When agents get published to Microsoft Teams and Microsoft 365 Copilot, Microsoft 365 admins can approve usage requests. These requests can be further scoped to a limited group for pilot testing/department usage or the full organization. Adding an AI gateway Observability tells you what your agents are doing, but how do you actually control them? This is where Azure API Management comes in. Once you have more than one agent, model deployment, or multiple teams consuming them, you need a single enforcement point between the agents and the resources they call. Adding Azure API Management in front of Microsoft Foundry gives you: Rate limiting and load balancing across regions and model deployments Consistent authentication and quota policies for model and tool traffic Usage tracking per team or cost center so you can accurately charge back to different departments Governed access to your custom and remote MCP servers Note: When choosing MCP servers, start with trusted, enterprise supported sources (GitHub, Microsoft, internally developed servers, etc.) that have documented security controls, enterprise authentication, clear ownership, and least-privilege permissions. Treat community MCP servers as untrusted until they have undergone a formal security review and are verified by organizations. Adding an AI gateway completes the governance picture. Azure Policy governs what can be deployed. RBAC governs who can build and manage agents. Observability provides evidence of how agents behave in production. The AI gateway extends governance into runtime, controlling how agents interact with models, tools, and external systems. Combined, these layers help organizations move beyond simply building agents to operating them responsibly at scale. Extend governance across the wider estate A few directions to take this further: MCP registry in Azure API Center – As MCP usage grows, you can create an approved inventory of MCP servers and APIs that can be used across an organization. Microsoft Agent 365 – Microsoft's enterprise control plane for AI agents. It gives every agent its own Microsoft Entra Agent ID and published Foundry agents sync to its registry automatically. This gives your IT team one place to run access reviews, lifecycle policies, and owner attestation across every agent in the tenant, including shadow agents discovered outside Foundry. Copilot Studio – when you add an MCP tool, you can point it at the API Management URL instead of the direct remote endpoint to gain additional observability through the gateway. GitHub Copilot – you can apply the same AI gateway-fronted MCP registry, applied to the developer side. Microsoft Purview – data classification, DLP, audit, and AI interaction governance across the wider estate. Where to go next Looking for a quick start? Turn on the three Azure Policy definitions above in audit mode against a non-production subscription and see what gets flagged that is out of compliance. Ready to design the end-to-end pattern? Take a look at this Cloud Adoption Framework guidance on AI governance and governing Azure platform services for AI. Want to go deeper on agent observability? Start with these articles around Application Insights integrations with Foundry: Use Insights in Microsoft Foundry and Monitor AI Agents with Application Insights Governing AI agents doesn't require starting from scratch. The identity, policy, monitoring, and cost controls you already use for the rest of your Azure estate can extend to AI workloads. Start with one layer, connect the next, and build a governance foundation that grows with your AI adoption.
Your Agents Need More Than a Place to Run
In architecture reviews with enterprise teams moving their first agentic applications toward production, I often hear the same plan: the team has containerized their agent and intends to run it on the managed Kubernetes cluster the organization already trusts. The reasoning is sensible, since the platform team knows the tooling and security has approved the network model, and for the first use case or two it is often the right call. Having watched this unfold in my years leading AgenticAI customer engineers and forward deployed engineers, and now helping customers reach production on Azure, I want to describe what happens next, before leaders commit rather than after. One thing to note is Azure supports multiple ways to build and operate agents. Foundry Agent Service provides an integrated managed runtime around your agent code. Azure Kubernetes Service supports teams that need Kubernetes-level control or want to extend an established platform, while Azure Container Apps provides managed container hosting. These services can work together. The decision is which capabilities and responsibilities best fit the workload. Where the cluster is the right answer A stateless retrieval application, a document extraction pipeline, or a classification job is a web service that happens to call a model, and a container platform runs web services well. A large retail customer of mine ran an invoice extraction agent on containers for over a year with almost no issues, and I never suggested they move it. The cluster remains the right home for several other situations as well. LLM invocations embedded inside existing microservices, event driven, or batch pipelines fit container platforms naturally. Genuine constraints such as air-gapped or sovereign environment, regions where a managed service is not yet offered are a good reason to run your own stack, and strong engineering team that already operates at that level can be a real asset. Even when the agents themselves move to a managed runtime, the tool servers, business APIs, and data services those agents call, often stay on your cluster, so they are complementary far more often than they are competitors. The problem is that these early wins can make agents seem like just another workload. Where the wheels come off The first failure is the state. A research agent that plans, searches, and synthesizes for ninety minutes is a long-running stateful process, while a Kubernetes pod is a disposable container the scheduler may restart at any time. A financial services team I worked with lost costly research run to a routine node upgrade, and responded as capable engineers do by building checkpointing, a durable store, and a resume mechanism. It worked, but they now owned a piece of infrastructure they had to keep correct as their agent's framework changed beneath it. A managed agent runtime absorbs this. Hosted agents in Foundry Agent Service, as one example, give every session a VM-isolated sandbox with a persistent file system, a durable state store that survives crashes and restarts and can hold checkpoints for frameworks such as LangGraph or Microsoft Agent Framework, and a resilient execution mode that recovers long-running work after a process interruption. The second is the human-in-the-loop. A commercial insurance customer’s claims agent needed sign-off from an adjuster and sometimes a second reviewer, with days between steps. Stopping an agent cleanly at the moment it needs a decision, holding its full session for four days without paying to keep it running, and resuming it correctly when the approval arrives is not something a container orchestrator gives out the box. The team built agent session suspend and resume, a queue, a durable state store, notifications, and an approval interface, and ended up with a small workflow engine nobody had planned to own. In hosted agents, an idle session is deprovisioned with its state persisted and restored onto fresh compute when the same session ID returns, so an agent waiting on an approval cost nothing while it waits, and sessions are retained for up to thirty days of inactivity. The approval experience remains yours to design, which is where your engineers' time should go. The third is multi-turn conversation, which quietly pushes teams into building their own context management system. The first version appends each turn to history, and within few turns the history outgrows the context window while cost and latency climb. So, the team adds truncation, then summarization, then retrieval of earlier turns, then per-user and per-tenant scoping, then expiry and deletion rules for privacy. An industrial customer's safety compliance assistant, with conversations stretching across days, followed exactly this path and ended up with a bespoke thread store and summarization pipeline nobody had budgeted for. Its first serious incident came when a summary silently dropped a compliance-relevant instruction. With the Responses protocol in hosted agents in Foundry, conversation history is a durable, platform-managed record keyed by a conversation ID and reachable from any channel, so the thread store is no longer yours to build, although deciding what to summarize or retrieve remains a design choice for your agent. The fourth is identity. On a cluster, the path of least resistance is a shared service account, and in one review a security architect asked which actions had been taken on behalf of which user, only to learn that the logs could not say. Hosted agents create a dedicated Microsoft Entra agent identity for each agent at deploy time, use on-behalf-of flows to act with the user's delegated permissions in interactive scenarios and the agent's own identity in autonomous ones, and keep the agent identifiable for audit in both cases. The fifth is per-user session isolation, which is the difference between an agent that serves many people and an agent that mixes them up. Agents read files, run code, and hold working data, and on a shared pod the default is that many users share a process, a file system, and often a cache. Giving every user session its own sandboxed environment and storage, so that one person's documents and intermediate results can never surface in another's, means engineering hard isolation boundaries and proving them to your security team. Hosted agents make a VM-isolated sandbox per session the default, and their durable state store can partition items per end user, so one store is safe to share across the users of a multitenant agent. And lastly, Evaluation and optimization are where the gap widens. The largest difference between teams that scale and those that stall is evaluation. Because agents are probabilistic and multi-step, staging tests often miss failures such as a wrong tool choice or a policy violation deep in a task. One customer’s agent passed every offline check but degraded unnoticed for weeks after a model update because evaluation stopped at release. Mature teams continuously evaluate production traces, combine automated judges with sampled human review, and red-team regularly.Once quality is measurable, teams can deliberately balance prompts, models, tools, latency, and cost. One team cut per-task cost by routing simple steps to smaller models after evaluation confirmed quality held. Self-built stacks require teams to assemble and maintain tracing, datasets, judges, and release gates. Hosted agents instead combines default OpenTelemetry traces with continuous evaluation, adversarial testing, and datasets generated from agent instructions. Staged closed-loop optimization uses those traces to improve instructions, tool descriptions, and model selection without extra plumbing. The real cost is the velocity gap Leaders usually expect me to quantify the initial build, and that is the smaller number. A team of four building the first agent often becomes ten or twelve within a year, and a growing share of them are maintaining a runtime for agents rather than building agents that serve the business. The enterprise has quietly created an internal agent infrastructure company in a field where the patterns for memory, tools, evaluation, and safety are rewritten every few months, while hyperscalers put hundreds of engineers on exactly this problem and ship at a cadence no single platform group can match. Foundry Agent Service, for instance, bundle content safety guardrails into the runtime and route outbound traffic through a customer virtual network, capabilities that platform teams otherwise assemble one integration at a time. This plumbing does not differentiate your organization, so the question is whether your scarcest engineers should spend years on it or on the workflows, data, and judgment only your company has. This is also why the technology companies held up as examples are a poor template. Many built their own runtimes because managed options did not yet exist and had large platform organizations to carry the load. Even they tend to invest in a custom runtime for the first handful of use cases and then migrate as managed services mature, because the maintenance burden compounds while the strategic value of owning the plumbing does not. What I would do as the leader I am not arguing against Kubernetes or for moving everything tomorrow. Comparable managed runtimes exist across the hyperscalers, and I use hosted agents as the running example only because it is the one I know best from the inside. Managed agent services are still maturing, some workloads have real data residency or customization needs, and abstraction always constrains something. What I am arguing for is a deliberate choice for each use case, guided by three questions: whether the agent must outlive a single request by running long, waiting on humans, or remembering across sessions whether it must act with its own identity, auditable delegation, and policy enforcement whether your team would be building anything a managed service already provides, and who will still maintain it in two years. Key takeaways Match the runtime to the workload. Containers on your existing cluster suit stateless, short-lived agents, while long-running, human-in-the-loop, and memory-dependent agents need capabilities your platform team was never hired to build. Count the hidden team. The real cost of self-hosting is the growing group of engineers maintaining state, identity, memory, guardrails, and tracing instead of solving business problems. Make evaluation continuous and connected to production traces. Pre-release testing alone will miss the drift and trajectory failures that hurt you, and optimization of quality, cost, and latency depends on that evaluation data. Follow the pattern of the leaders, not their early architecture. Companies that built custom runtimes did so before managed options matured, and most of them move toward managed services as those options improve. A practical place to begin is your roadmap for the next twelve months. Sort each use case into stateless and short-lived or long-running and human-dependent, and for every agent in the second group put a price on the engineers who would maintain the runtime rather than the business logic. I would like to hear how you have drawn this line, and where a managed service was not yet ready for something you needed.Beyond the Trace: The Science of Insight Quality
Co-authors and reviewers: Morteza Ziyadi, Hanchi Wang, Han Che, Billy Hu, Sean Gayler, Nishal Dsilva, Avinav Jami, Ankit Singhal, Augustus Arthur TL;DR. Insights in Foundry turns recurring agent behavior into evidence-linked findings developers can review and act on. We evaluate trace linkage, finding quality, and detection of known issues using labeled benchmarks, LLM-judge assessment, and controlled end-to-end tests. What is an Insight? An Insight is a reviewable finding about recurring agent behavior. It brings together an explanation, supporting trace evidence, and a possible next step, helping developers investigate a pattern rather than inspect each execution in isolation. Depending on the available evidence and supported configuration, an Insight can include: Part of an Insight What it gives the reviewer Title and description The recurring behavior and an evidence-based explanation of a possible cause. Linked traces The broader group of traces associated with the finding. Highlighted traces Representative examples to open and check against the explanation. Category, severity, and status Context for triaging the finding, not a substitute for a risk assessment. Agent version and recency Which version is represented and when the finding was created. Proposed action or fix An investigation or improvement path; concrete proposals depend on the configuration. Figure 1. An Insight detail view in Microsoft Foundry, cropped from public Microsoft Learn documentation. The description, evidence, and proposed fix support human review. This UI illustration is not a benchmark result; the source link provides a full-size view. UI source and field definitions: Insights in Foundry documentation. From trace search to an actionable review queue Production agents can generate thousands of traces containing model calls, tool calls, latency, token usage, errors, and final responses. Observability shows what happened, while evaluations test criteria a team already knows to measure. The harder problem is discovering repeated behavior that the team did not know to predefine. Insights in Foundry analyzes traces from the Application Insights resource connected to a Foundry project and organizes recurring behavior into reviewable Insights. Depending on the evidence and supported configuration, an Insight can include representative traces, affected agent versions, severity, an explanation of a possible cause, and a recommended next action. Concrete prompt or code proposals are available only for supported agent types and configurations; other Insights provide general investigation or remediation guidance. Insights in Foundry is now available in public preview. It analyzes production traces to surface recurring behaviors and regressions with supporting evidence and suggested areas for investigation or improvement. Developers remain in control: review the cited traces and validate proposed changes through normal evaluation and deployment practices. During public preview, we will continue expanding the experience based on customer feedback. To make this concrete, Figure 2 maps 220 inputs from TraceElephant, a public agent-trace benchmark, to seven generated Insights. Figure 3 examines one linked trace. Figure 2. Recorded links from 220 public TraceElephant inputs to seven generated Insights in the September 17 benchmark. The 80-input finding includes the looping trace examined in Figure 3. Titles are editorial paraphrases; the layout is editorial, not a product screenshot or a validation of each diagnosis. Figure 3. A public TraceElephant trace contains 54 model calls; steps 26, 30, 34, 38, 42, 46, and 50 repeat an identical extraction plan. The September 17 benchmark run of the production pipeline links this trace to a finding about missing progress-aware termination. Wording is paraphrased; this is an illustration, not product UI. The proposed intervention still needs evaluation. Three complementary ways to evaluate Insight quality Insight quality involves which inputs a finding links, how well it explains the evidence, and whether it identifies known issues in controlled tests. We combine labeled trace evaluation, unlabeled finding assessment, and controlled end-to-end testing. These are complementary sources of evidence, not mutually exclusive dataset categories. Evidence Question Assessment / results Labeled trace evaluation Do Insights link inputs that reference annotations mark as failing? Reference labels; trace precision and trace recall (%). Unlabeled finding assessment Are generated findings grounded and useful? Eight-dimension LLM judge; mean score (1-5). Controlled end-to-end tests Do Insights identify known injected issues? Known issues and healthy baselines; detections, misses, unsupported findings, duplicates, and unscored cases. Here, 'unlabeled' means reference failure labels are not used for the evaluation; the sources can still contain annotations or reference answers. Public versus internal describes where data comes from, not how it is evaluated. LLM judges can also assess findings from labeled or controlled scenarios. Labeled evaluation: Do Insights link the right traces? Trace precision: among unique benchmark inputs linked by at least one generated Insight, the fraction marked as failing. This measures discrimination only when a dataset includes healthy inputs, and it does not validate the diagnosis. Trace recall: the percentage of failure-labeled benchmark inputs linked by at least one generated Insight. It does not measure how many distinct failure modes were discovered. We evaluated the production pipeline on the same 746 benchmark inputs (653 failure-labeled) on September 15, 16, and 17, 2026. Repeating these inputs does not create 2,238 distinct examples. The six benchmark corpus slices come from public agent-trace research datasets with human reference annotations: AgentRx (Tau-bench retail), AgentRx Magentic-One, TRAIL, AgentErrorBench, TraceElephant, and the MAST-Data human subset. AgentErrorBench is the annotated failure-trajectory dataset released with AgentDebug, a framework for detecting and recovering from agent failures. The two AgentRx slices come from the same public release. The benchmark uses selected and normalized inputs from these releases, not a representative sample of customer production traffic. For each dataset, the chart shows daily generated-Insight counts and trace recall. The table reports equal-weight means of daily trace precision and recall. These datasets support ongoing development; results depend on the dataset and model used. Labeled results Figure 4. Generated Insight count versus trace recall for all six public corpus slices. Each point is one daily run; labels retain values where markers overlap. Recall measures linkage to failure-labeled inputs, not diagnosis correctness. Three-day mean precision and recall follow in the table. Dataset Inputs/day Failure-labeled/day Linked/day range Mean trace precision Mean trace recall AgentRx (Tau-bench retail) 102 29 43-45 37.8% 57.5% AgentRx Magentic-One 58 44 54-57 75.3% 94.7% TRAIL 148 143 139-140 96.9% 94.6% AgentErrorBench 200 200 194-196 100.0%** 97.3% TraceElephant 220 220 218-220 100.0%** 99.7% MAST-Data (human subset) 18 17 14-16 93.3% 82.4% ** When every input is failure-labeled, 100% trace precision cannot measure false-positive control. What the labeled results show High precision can reflect the corpus base rate. AgentErrorBench and TraceElephant contain only failure-labeled inputs, so their measured precision is 100% by construction. TRAIL, AgentRx Magentic-One, and the MAST subset also contain mostly failures; their precision values are close to their underlying failure-label prevalence. Read precision beside trace recall and each dataset's label prevalence rather than as a standalone quality score. AgentRx (Tau-bench retail) is the clearest mixed-traffic stress test. Only 29 of 102 inputs were failure-labeled. Across the three days, trace precision was 43.2%, 37.8%, and 32.6%, averaging 37.8%; trace recall was 65.5%, 58.6%, and 48.3%, averaging 57.5%. Daily linked counts were 44, 45, and 43, including 19, 17, and 14 labeled failures respectively. These daily rates are averaged before rounding. Some linked inputs without failure labels could contain issues outside the reference labels, such as cost or latency, but this benchmark cannot confirm that. Trace precision and trace recall measure whether Insights link inputs that reference annotations mark as failing. They do not establish whether an explanation is correct or a proposed action is useful. Our unlabeled evaluation examines those qualities using an eight-dimension LLM judge. Unlabeled evaluation: Are the findings grounded and useful? For this evaluation, a separate LLM judge scores generated Insights against the supplied input evidence, without using reference failure labels to compute precision or recall. These scores are automated assessments produced by an LLM judge. They are not human ratings and do not establish that a proposed action will improve the agent. This evaluation covers PUPA, FailSafeQA, tau2-bench, synthetic scenarios, and an internal production dataset for the monitoring dashboard agent compiled from Microsoft employee usage. These sources are normalized into benchmark inputs; not every input is a complete execution trace. The eight-dimension judge rubric Each generated Insight receives a score from 1 to 5 on eight dimensions. The rubric separates the quality of the explanation, the importance of the issue, and the usefulness of the suggested action: Dimension Question the judge considers Actionability Does the Insight tell a developer what to investigate, evaluate, or change next? Specificity Does it name concrete behavior, tools, or prompt elements rather than offer generic advice? Novelty Does it surface a pattern beyond an obvious dashboard signal? This is an estimate, not a measurement of what a team already knows. Correctness Is the claim supported by the supplied evidence? Severity calibration Does the assigned severity match the evidenced impact? Impact How consequential does the underlying issue appear, rather than just how well it is described? Fix specificity Does the proposed fix identify an exact asset or behavior to change? Fix applicability Is that proposed change plausibly on target for the issue? For example, a concrete next step earns a stronger actionability rating than generic advice. Severity calibration and impact are deliberately separate: a minor issue can have a well-calibrated severity label without being high-impact. Fix applicability is judged from the proposal and evidence; the fix is not executed as part of this scoring. Unlabeled results Results compiled from two recent benchmark runs cover 4,496 inputs and 40 generated Insights. Every emitted Insight in these selected snapshots was scored with the same judge rubric. Each Insight's overall score is the mean of its eight dimension scores; the table then averages those overall scores within each dataset. We do not combine datasets into one quality score. Dataset Inputs Insights scored Mean judge score (1-5) PUPA 901 5 3.53 FailSafeQA 1,101 11 3.57 Monitoring dashboard agent (internal) 1,342 4 3.88 tau2-bench 1,112 16 2.84 Synthetic scenarios 40 4 3.81 These are per-dataset snapshots, not a controlled comparison of dataset difficulty or pipeline versions. Small Insight counts and grading choices, such as grouped versus individual judging, make these protocol-specific diagnostics rather than stable absolute ratings. Automated judge scores are not precision or recall, human ratings, or measured improvements after applying a fix. The dimension-level scores make the weak spots more concrete. Fix specificity was the lowest-scoring dimension in four of the five snapshots, with means from 2.00 to 2.64. For tau2-bench, correctness was lowest at 2.25: a signal to review whether claims are supported by the supplied evidence. Making a next step more concrete and grounding a diagnosis more carefully are different improvement targets. What the public findings look like The examples below connect generated findings to concrete input evidence: an arithmetic inconsistency, a question/answer mismatch, and a payment allocation that exceeds the available balance. They are selected illustrations, not a representative sample of the 40 scored Insights. Figure 5. Three illustrative findings, with one checked input per finding. Titles and evidence summaries are editorial paraphrases. PUPA is a normalized QA record; FailSafeQA normalization pairs a perturbed question with the original answer, creating the displayed mismatch. Neither is a new agent execution. Tau2-bench shows a recorded tool interaction. These examples are not a representative quality sample. Public input sources: PUPA; FailSafeQA; tau2-bench. Controlled end-to-end tests: Can Insights find known issues? Dataset benchmarks are complemented by tests with healthy baseline agents and versions containing predefined defects. The framework generates traffic, runs Insights on the resulting traces, and compares the findings with the known issues and supporting evidence. It separately tracks detections, misses, unsupported findings, and duplicates. Cases with incomplete evidence remain unscored rather than counting as passes or misses. The September 23 hosted-agent report, used here as a concrete illustration, covered finance, travel, and support-ticket scenarios. Of 12 expected issues, 11 were scorable and 10 were detected. All three scorable healthy baselines had no confirmed unsupported findings. The figure shows the full outcome counts, including the missed and unscored cases. This end-to-end framework is separate from the 40-input synthetic-scenarios dataset in the unlabeled results. Figure 6. Controlled tests exercise healthy baselines and known defects through the Insights pipeline. The September 23 report records 10 detections among 11 scorable expected issues, one additional unscored issue, and no confirmed unsupported findings on three scorable baselines. No confirmed unsupported findings (noise) or duplicate findings were reported in the evaluated subset. This illustrative snapshot is not a representative product-wide rate. A broader view of quality Trace selection and rubric review are part of a broader quality framework. Category agreement is an exploratory benchmark diagnostic, not an assessment of the portal's category labels. Controlled scenarios also check for unsupported or duplicate findings. Figure 7. Five complementary quality questions. The labeled results above measure trace precision and trace recall; the unlabeled results assess findings with an LLM judge. Category agreement is an internal diagnostic, while controlled scenarios check noise and duplication. How we track quality over time Labeled benchmarks, unlabeled assessment, and controlled tests feed recurring quality reports and human investigation. Test inventories and assessment conditions can change, so daily scores are not automatically a comparable improvement trend. The reported results do not establish longitudinal recurrence or deduplication performance across runs. In the portal, Give feedback lets users report incorrect findings, categories, severity, grouping, or duplication. Start with scenarios whose expected failures and healthy controls are known. Track detection, trace precision, categorization, severity, distinctness, evidence grounding, and action quality instead of relying on one score. Review changes to data, models, prompts, and evaluation contracts as changes to the measurement system. Investigate weak results and newly reported failure patterns rather than tuning only for an aggregate. Keep human review in the loop, because benchmark labels and automated graders cannot determine business impact or remediation correctness. How practitioners should interpret an Insight Validate each Insight before acting on it. Microsoft Learn recommends this review sequence: Confirm the affected workflow, agent version, category, severity, and time. Open the highlighted traces and verify that the cited behavior is present. Compare problematic examples with healthy traces to test whether the grouping and likely cause are plausible. Decide whether the issue belongs to the agent, a tool, a model endpoint, a data source, or the platform. Turn confirmed recurring behavior into evaluation coverage, an optimization objective, owner routing, or a monitored no-action decision. An empty result does not prove that an agent is healthy, and a large linked-trace count does not prove business impact. Quality depends on representative traces, complete telemetry, a supported analysis model, and careful human validation. Get started Start with the Insights in Foundry documentation for prerequisites, portal steps, evidence-review guidance, SDK examples, pricing considerations, and preview limitations. Prerequisites include a connected Application Insights resource, recent representative traces, a supported GPT-5-or-newer Judge model deployment, and the required role assignments. Insight generation, including scheduled runs, uses your model deployment and can incur model charges. See the current documentation for the latest supported agents, models, regions, limits, pricing, and UI guidance. Python samples: on-demand Insights and scheduled Insights. Closing thoughts Agent quality work often starts with a simple question: what is repeatedly going wrong that we did not know to test? Insights in Foundry is designed to help teams answer that question with evidence-linked findings and a reviewable next step. Measuring those findings requires explicit definitions, diverse data, honest treatment of weak results, and human judgment where automated metrics stop. Labeled trace evaluation, unlabeled rubric assessment, and controlled end-to-end tests answer complementary questions. Newly observed patterns can then inform the next round of evaluation and review.How to make AI responses faster on Microsoft Foundry: lessons from 2,040 measurements
A reproducible Microsoft Foundry performance study covering prompt caching, multimodal input, tool orchestration, MCP lifecycle, Toolbox tool search, and Priority Processing - with correctness and reliability measured alongside latency.Applying Site Reliability Engineering to Autonomous AI Agents
If you practice SRE, you already have a mental model for running reliable production systems. You define SLOs. You track error budgets. You use circuit breakers to stop cascading failures. You run chaos experiments to find weaknesses before customers do. You treat every operational decision as a tradeoff between reliability and velocity. That mental model transfers directly to AI agents. It just needs four new ideas. In the Agent Governance Toolkit: Architecture Deep Dive, Policy Engines, Trust, and SRE for AI Agents, we covered Agent SRE briefly as one of AGT's nine packages: SLOs, error budgets, circuit breakers, chaos engineering, and progressive delivery, adapted from the patterns your SRE team already applies to microservices. Several teams asked for the full story. This is it. Agent SRE is one of the more novel parts of the toolkit. The policy engine, zero-trust identity, and execution sandboxing have clear analogs in existing security practice. Agent SRE explores newer ground. Established patterns for defining SLOs for AI agent behavior, building chaos experiments for LLM provider failures, or applying error budgets to agent autonomy are still emerging across the industry. We built these capabilities because running agents in production without them is the equivalent of running a fleet of microservices without circuit breakers, health checks, or an on-call runbook. This post is for SRE teams, platform engineers, and anyone responsible for running AI agents in production. You do not need to be an AI specialist. If you know what a burn rate is, you are ready for this. The Problem: Agents Fail in Ways Your Existing SRE Tooling Cannot See When a service fails, your observability stack tells you: latency went up, error rate crossed the SLO threshold, the circuit breaker opened. You page the on-call engineer. They look at traces and find the slow database query. When an AI agent fails, your observability stack is silent. The agent returned HTTP 200. Latency was normal. Error rate was zero. But the agent quietly approved a transaction it was not authorized to approve, hallucinated a database path and wrote to the wrong table, or got stuck in a reasoning loop that consumed $800 of LLM API budget before anyone noticed. These are not infrastructure failures. They are behavioral failures. And they are invisible to monitoring tools built for stateless, deterministic services, because those tools only watch for crashes and timeouts. They do not watch for wrong behavior. This gap is the problem Agent SRE was designed to solve. The solution borrows everything from the SRE playbook and adds one concept that extends it: the Safety SLI. The Safety SLI: A New Reliability Dimension Traditional SLIs measure system behavior from the user's perspective: latency, availability, error rate, throughput. They answer: did the service respond correctly? For AI agents, correctness is not enough. An agent that responds correctly but acts outside its authorized scope has not succeeded. It has failed in a way that none of your existing SLIs can detect. The Safety SLI answers a different question: did the agent act within policy? from agent_sre import SLO, ErrorBudget from agent_sre.slo.indicators import PolicyCompliance # Define a safety SLO: 99% of agent actions must comply with policy safety_slo = SLO( name="safety-compliance", indicators=[ PolicyCompliance( target=0.99, window="7d", ), ], error_budget=ErrorBudget( total=0.01, # 1% budget (1 - 0.99 target) window_seconds=2592000, # 30-day window burn_rate_alert=2.0, # warn at 2x sustainable rate burn_rate_critical=5.0, # page at 5x sustainable rate ), ) When an agent's policy compliance rate drops below 99%, the error budget starts burning. The ErrorBudget tracks consumption automatically and exposes burn rate alerts through its firing_alerts() method. When the budget is exhausted, the configured exhaustion_action determines the system response: from agent_sre.slo.objectives import ExhaustionAction # Configure what happens when error budget is exhausted safety_slo = SLO( name="safety-compliance", indicators=[PolicyCompliance(target=0.99, window="7d")], error_budget=ErrorBudget( total=0.01, window_seconds=2592000, burn_rate_alert=2.0, # fires at 2x sustainable burn rate burn_rate_critical=5.0, # fires at 5x sustainable burn rate exhaustion_action=ExhaustionAction.CIRCUIT_BREAK, # suspend agent when budget is gone ), ) # In your monitoring loop, check for firing alerts alerts = safety_slo.error_budget.firing_alerts() for alert in alerts: print(f"Alert firing: {alert.name} (severity: {alert.severity})") # Check budget status print(f"Budget remaining: {safety_slo.error_budget.remaining_percent:.1f}%") print(f"Current burn rate: {safety_slo.error_budget.burn_rate():.2f}x") print(f"Exhausted: {safety_slo.error_budget.is_exhausted}") This is the governance dial from the other direction. The error budget is not just a metric: it is the mechanism that drives agent autonomy decisions. An agent with a clean 30-day safety record earns autonomy. An agent whose budget is burning at 5x the sustainable rate triggers a critical alert, and when the budget is exhausted, the exhaustion_action fires: ALERT, THROTTLE, FREEZE_DEPLOYMENTS, or CIRCUIT_BREAK. The graduated response mirrors what SRE teams already do with service SLOs, applied to agent behavior. There are multiple SLI dimensions built into Agent SRE. Safety SLIs and Performance SLIs track different aspects of the same agent: SLI Type What It Measures Target Pattern When Budget Burns Safety SLI PolicyCompliance -- fraction of actions within authorized scope >= 99% Restrict capabilities, increase human oversight Performance SLI TaskSuccessRate, ResponseLatency, CostPerTask Configurable per workload Alert, throttle, or circuit-break LLM provider Additional built-in indicators include ToolCallAccuracy, DelegationChainDepth, HallucinationRate, and CalibrationDeltaSLI. Both SLOs feed into the same error budget dashboard. An agent can have excellent performance but a degrading safety record, or perfect safety compliance and terrible cost efficiency. You need both dimensions to understand whether an agent is production-ready. Circuit Breakers: Governing Agent Failure Modes That Don't Exist in Microservices Circuit breakers for services protect against one failure mode: a backend that is slow or unreachable. The pattern is CLOSED -> OPEN -> HALF_OPEN. You know it well. Agent SRE implements the same state machine for failure modes that are specific to autonomous reasoning systems and do not exist in traditional microservice architectures: from agent_sre.cascade.circuit_breaker import CircuitBreakerConfig, CircuitBreaker from agent_sre.chaos.engine import FaultType config = CircuitBreakerConfig( failure_threshold=5, # Open after 5 failures in the window recovery_timeout_seconds=60, # Stay OPEN for 60s before HALF_OPEN half_open_max_calls=3, # Allow 3 probes in HALF_OPEN ) breaker = CircuitBreaker(agent_id="analyst-agent-001", config=config) # Failure modes tracked by the circuit breaker: tracked_faults = [ FaultType.POLICY_BYPASS, # Agent exceeds authorized scope FaultType.ERROR_INJECTION, # Upstream model API fails FaultType.TIMEOUT_INJECTION, # Tool calls exceed time budget FaultType.TRUST_PERTURBATION, # Agent trust score falls below threshold FaultType.DEADLOCK_INJECTION, # Agent stuck in iterative reasoning ] Each failure mode has different circuit-breaking semantics: Failure Mode What Triggers It Circuit-Break Behavior Policy bypass Action denied by policy engine Count toward threshold; log with full context LLM provider error HTTP 5xx from model API Immediately open; route to fallback model if configured Tool timeout Tool call exceeds timeout_ms Count toward threshold; cancel in-flight call Trust score degradation Agent trust score drops below configured floor Open; escalate to Ring 3 (untrusted) until score recovers Reasoning loop / deadlock Token or iteration count exceeds budget Open; trigger human review before resuming The reasoning loop breaker deserves attention. A microservice cannot get stuck reasoning. An AI agent absolutely can, and when it does, the failure is not an error code: it is an agent that keeps calling tools, consuming tokens, and generating audit events indefinitely. The circuit breaker detects this pattern from the iteration count and token budget and terminates the loop: # Reasoning loop detection configuration loop_detection_config = { "max_iterations": 15, # Hard stop after 15 reasoning steps "max_tokens_per_session": 50000, # Hard stop on token consumption "repetition_threshold": 0.85, # Stop if >85% of recent actions repeat prior ones "on_detection": "circuit_break_and_escalate", } The state machine behaves identically to what you know from Hystrix or Resilience4j. What changes is the definition of "failure." CLOSED (serving) | | failure_threshold crossed for any tracked fault v OPEN (rejecting -- agent action denied, fallback or human-in-loop fires) | | recovery_timeout expires v HALF_OPEN (probe -- limited requests allowed through) | |-- success_threshold met --> CLOSED |-- any failure --> OPEN (reset timeout) Chaos Engineering for Agents: Fault Injection for Autonomous Systems The only way to know if your agent system is resilient is to break it intentionally. Traditional chaos engineering targets infrastructure: kill a pod, inject network latency, saturate a disk. Agent chaos engineering targets the failure modes specific to autonomous reasoning systems. Agent SRE ships fault injection templates that cover the failure modes teams consistently underestimate until they hit production: from agent_sre.chaos.engine import ChaosExperiment, Fault, FaultType # Experiment 1: LLM provider degrades -- model returns valid responses but with # increased latency and occasional malformed outputs experiment = ChaosExperiment( name="llm-degradation-resilience", target_agent="analyst-agent-001", description="Test agent behavior under degraded LLM provider", faults=[ Fault.latency_injection(target="llm-provider", delay_ms=8000), Fault.error_injection(target="llm-provider", rate=0.05), ], duration_seconds=300, ) # Experiment 2: Trust score manipulation -- simulates an agent receiving # messages from a peer with a spoofed trust score trust_experiment = ChaosExperiment( name="trust-manipulation-resilience", target_agent="orchestrator-001", faults=[ Fault( fault_type=FaultType.TRUST_PERTURBATION, target="did:mesh:orchestrator-001", params={"spoofed_score": 950}, ), ], duration_seconds=120, ) # Experiment 3: Tool timeout cascade -- multiple tools time out simultaneously, # testing whether the agent abandons gracefully or enters a reasoning loop cascade_experiment = ChaosExperiment( name="tool-timeout-cascade", target_agent="analyst-agent-001", faults=[ Fault.timeout_injection(target="database.read", delay_ms=30000), Fault.timeout_injection(target="api.call", delay_ms=30000), ], duration_seconds=180, ) # Run the experiment experiment.start() # ... inject faults during agent execution ... resilience = experiment.calculate_resilience( baseline_success_rate=0.95, experiment_success_rate=0.87, recovery_time_ms=48000, ) experiment.complete(resilience=resilience) print(f"Resilience score: {resilience.overall}/100 -- {'PASSED' if resilience.passed else 'FAILED'}") Additional fault types built into the chaos engine cover: prompt injection attempts, privilege escalation, data exfiltration attempts, identity spoofing, deadlock injection, and contradictory instruction scenarios. Each maps to a FaultType enum value and can be composed into multi-fault experiments. Important: The chaos engine records that a fault was injected and triggers the governance response pipeline. Actual infrastructure-level fault injection (network partition, process kill) should be implemented using your existing chaos tooling (Chaos Mesh, Gremlin, Azure Chaos Studio, or similar). Agent SRE governs the agent's behavioral response to faults; it does not own infrastructure manipulation. These two layers are designed to compose. Each chaos experiment produces a structured resilience score via calculate_resilience(), which compares baseline and experiment success rates. A score of 90+ with passed=True means the agent maintained at least 90% of its baseline performance under fault conditions. Teams use this to set minimum resilience thresholds for production readiness. Replay Debugging: Reproduce Behavioral Failures Exactly Infrastructure incidents are reproducible because infrastructure is deterministic. AI agent incidents are hard to reproduce because agent behavior depends on model state, context window content, and the sequence of tool call results, none of which are preserved by default after a session ends. Agent SRE's replay engine records every agent session as a replayable artifact: the full trace at each step, every tool call with its inputs and outputs, every policy evaluation with its decision, and every trust score at the time of each inter-agent message. from agent_sre.replay.capture import TraceStore from agent_sre.replay.engine import ReplayEngine, ReplayMode # Traces are captured automatically when SRE tracing is active store = TraceStore( backend="azure_blob", retention_days=30, ) # When an incident occurs, replay the session exactly engine = ReplayEngine(store=store) # Full replay: re-run the session against the same recorded inputs # Uses recorded tool outputs -- no live tool calls -- so replay is deterministic result = await engine.replay( trace_id="trace_2026_05_a7f3b2", mode=ReplayMode.FULL, ) for step in result.steps: print(f"Step {step.index}: {step.action} -> {step.decision}") # Divergence analysis: replay with a policy change applied # Shows exactly which actions would have been blocked under the new policy diff_result = await engine.diff( trace_id="trace_2026_05_a7f3b2", policy_override="policies/stricter-v2.yaml", ) for diff in diff_result.diffs: if diff.description: print(f"Step {diff.span_name}: was {diff.original}, " f"would be {diff.replayed} under new policy") The divergence analysis is the feature teams use most. When a policy change is proposed, you replay recent production traces against the new policy to see how many actions would have been blocked, which sessions would have failed, and what the error budget impact would have been. Policy changes stop being guesswork. Progressive Delivery: Safely Rolling Out New Agent Capabilities When you ship a new service version, you do not send it to all traffic at once. You use canary deployments, feature flags, or traffic splitting. You watch the SLOs. If they degrade, you roll back. Agent SRE brings the same discipline to agent capability rollout. When you expand an agent's authorized scope, giving it write access it did not have, connecting it to a new tool, or raising its trust floor, you do not expand to the full fleet immediately. You expand progressively, with automated SLO gates controlling each stage. from agent_sre.delivery.rollout import ( AnalysisCriterion, CanaryRollout, RollbackCondition, RolloutStep, ) rollout = CanaryRollout( name="database-write-capability", steps=[ RolloutStep( name="canary", weight=0.05, # 5% of agents get the new capability duration_seconds=86400, # 24 hours analysis=[ AnalysisCriterion(metric="safety_sli", threshold=0.995), AnalysisCriterion(metric="performance_sli", threshold=0.90), AnalysisCriterion( metric="error_budget_consumed", threshold=0.10, comparator="lte", # canary can burn at most 10% ), ], ), RolloutStep( name="early-adopters", weight=0.25, # 25% traffic duration_seconds=172800, # 48 hours analysis=[ AnalysisCriterion(metric="safety_sli", threshold=0.990), AnalysisCriterion(metric="performance_sli", threshold=0.88), ], ), RolloutStep( name="general-availability", weight=1.0, # 100% traffic duration_seconds=604800, # 1 week of full observation analysis=[ AnalysisCriterion(metric="safety_sli", threshold=0.990), AnalysisCriterion(metric="performance_sli", threshold=0.85), ], ), ], rollback_conditions=[ RollbackCondition(metric="safety_sli", threshold=0.95, comparator="lte"), ], ) # Start the rollout -- SLO gates evaluate at each step rollout.start() # Advance to next step when analysis criteria pass if rollout.advance(): print(f"Advanced to step: {rollout.current_step.name}") print(f"Progress: {rollout.progress_percent:.0f}%") The SLO gate at each step is the same mechanism as a CI/CD quality gate, but measured on live production behavior rather than test results. An agent capability that degrades the safety SLI during canary does not promote to the next step. If a RollbackCondition fires, the rollout rolls back automatically. This is the mechanism that makes it operationally safe to expand agent autonomy: every expansion is measurable, every measurement gates the next expansion, and rollback is automatic. Health Checks and Backpressure Traditional health checks answer: is the service alive? For agents, alive is not enough. A healthy agent is one that is alive, operating within policy, consuming resources within budget, and maintaining a trust score above the Ring threshold it was assigned. # Agent health check covering multiple dimensions health = await agent_health_check( agent_id="analyst-agent-001", dimensions=[ "liveness", # Is the agent process running? "policy_compliance", # Is safety SLI above threshold? "trust_score", # Is trust score above Ring floor? "resource_budget", # Is token/API spend within limits? "tool_availability", # Are the tools the agent needs reachable? ], ) # health.status: "healthy" | "degraded" | "unhealthy" # health.dimensions: per-dimension pass/fail with values # health.recommended_action: "none" | "restrict" | "suspend" | "terminate" When health checks report degradation, backpressure controls engage before the circuit breaker opens. Backpressure is the earlier, softer response: accept fewer concurrent tasks, reject low-priority work, drain in-flight tasks gracefully before the situation escalates. # Backpressure configuration backpressure_config = { "backpressure_threshold": 0.80, # Engage when resource utilization > 80% "max_concurrent": 5, # Hard cap on simultaneous agent tasks "priority_shedding": True, # Drop low-priority tasks first "drain_timeout_seconds": 30, # Allow in-flight tasks to complete } The ordering matters: backpressure first, then circuit breaker, then suspension. Each stage is recoverable. Each stage preserves more agent state than the next. The SRE principle of graduated response applies to agents exactly as it applies to services. Observability: Governance Metrics Flow Into Your Existing Stack Agent SRE does not ask you to adopt a new observability platform. Governance metrics are exported through the same adapters your infrastructure monitoring already uses, including OpenTelemetry, Prometheus, Datadog, and others. from agent_sre.tracing.exporters import configure_exporters configure_exporters( backends=[ {"type": "prometheus", "endpoint": "http://prometheus:9090"}, {"type": "opentelemetry", "endpoint": "http://otel-collector:4317"}, ], include_metrics=[ "slo.safety_sli", # Per-agent safety compliance rate "slo.error_budget_remaining", # Error budget in percentage "slo.burn_rate", # Current burn rate vs sustainable "circuit_breaker.state", # CLOSED / OPEN / HALF_OPEN "circuit_breaker.failure_count", "trust_score.current", # Agent trust score (0-1000) "trust_score.ring", # Current execution ring "chaos.experiments_run", # Chaos experiment telemetry "health.status", # Aggregate health status "backpressure.load", # Current load vs threshold ], ) Key governance metrics available in your existing dashboards: Metric What It Tells You Alert Condition slo.safety_sli Fraction of agent actions within policy < 0.99 slo.burn_rate Rate at which error budget is consumed > 2.0 (warn), > 5.0 (page) slo.error_budget_remaining Budget left for the SLO window < 20% circuit_breaker.state Current breaker state per agent OPEN or HALF_OPEN trust_score.ring Execution ring (privilege level) Ring 3 (untrusted) health.status Aggregate health across all dimensions degraded or unhealthy If you are already running Grafana dashboards for your services, a governance dashboard for your agent fleet is a new data source and a new set of panels, not a new monitoring stack. The SRE Mental Model for Agents: Four New Concepts Everything in Agent SRE is built on the SRE mental model you already have, extended with four concepts that adapt traditional reliability thinking for autonomous systems: Traditional SRE Agent SRE Equivalent What Changes Latency SLI Safety SLI Correctness of *action*, not speed of *response* Error budget Autonomy budget Burns on policy violations, not just errors Circuit breaker Behavioral circuit breaker Opens on wrong *behavior*, not just failure codes Canary deployment Capability rollout Rolls out *scope*, not just code The governance insight is that error budgets work in both directions for agents. A service's error budget only decreases. An agent's autonomy is also a budget: it grows when the safety SLI is strong and shrinks when it degrades. The error budget mechanism becomes the operational mechanism for expanding and contracting agent autonomy in response to evidence, which is exactly what regulated industries and risk-averse enterprise teams need before they will trust an autonomous agent with consequential actions. Getting Started with Agent SRE pip install agent-sre A minimal Agent SRE integration requires three things: a safety SLO definition, a circuit breaker, and a health check. The progressive delivery and chaos engineering features layer on top when you are ready for them. from agent_sre import SLO, ErrorBudget from agent_sre.slo.indicators import TaskSuccessRate from agent_sre.cascade.circuit_breaker import CircuitBreakerConfig, CircuitBreaker # Step 1: Define your safety SLO slo = SLO( name="production-safety", indicators=[TaskSuccessRate(target=0.99, window="24h")], error_budget=ErrorBudget(total=0.01, burn_rate_alert=2.0, burn_rate_critical=5.0), ) # Step 2: Configure a circuit breaker breaker_config = CircuitBreakerConfig( failure_threshold=5, recovery_timeout_seconds=60, half_open_max_calls=3, ) breaker = CircuitBreaker(agent_id="my-agent", config=breaker_config) # Step 3: Wire into your existing agent loop async def governed_agent_loop(agent, task): # Check health first if not await agent_is_healthy(agent.id): return {"error": "agent suspended", "reason": "health check failed"} # Run within circuit breaker protection async with breaker: result = await agent.run(task) slo.record_event(good=result.policy_compliant) return result The quickstart in the repository walks through a complete setup with safety SLOs, circuit breakers, and a Prometheus dashboard export in under 50 lines. Why This Matters Most AI observability tools today focus on what you might call model quality: hallucination rate, latency, token cost, task completion. These are useful metrics. They are not SRE metrics. They do not answer whether the agent acted within its authorized scope, whether its behavioral error budget is burning at a dangerous rate, or whether it would survive the LLM provider going down. Agent SRE answers those questions using the operational vocabulary that SRE teams already understand: SLOs, error budgets, circuit breakers, chaos experiments, and health checks. The goal is not to replace your observability stack. It is to make agent governance visible inside it. The reliability of an autonomous agent is not a property of the model. It is a property of the governance infrastructure around it. Agent SRE is that infrastructure. Resources GitHub: github.com/microsoft/agent-governance-toolkit Install: pip install agent-sre Tutorials: 40+ tutorials including dedicated Agent SRE walkthroughs for SLO setup, chaos experiments, and progressive delivery Architecture reference: ARCHITECTURE.md OWASP compliance mapping: OWASP-COMPLIANCE.md -- Agent SRE addresses ASI-08 (Cascading Failures) directly through circuit breakers and SLO-based fault detection Part 1 -- Runtime governance: Policy engines, trust, and SRE overview Part 2 -- Shift-left governance: Catching violations before production Part 3 -- Post-hoc accountability: After the agent acts The Agent Governance Toolkit is an open-source project released under the MIT License. All features described in this post are available in the public repository. The `agent-sre` package is currently in public preview; APIs may change before general availability. Questions about Agent SRE in your environment? Open an issue at aka.ms/agent-governance-toolkit or start a discussion in the comments below.Building Production-Ready AI Agents in Microsoft Foundry: 10 Lessons Learned
A Real-World Scenario Imagine a customer support agent that answers invoice questions. During testing, everything works perfectly. But in production: The Finance API occasionally times out. The knowledge base contains outdated information. Tool calls fail during peak traffic. Token consumption rises unexpectedly. The result isn't a broken AI model. It's an unreliable system. This is where Microsoft Foundry becomes critical. Building production-ready agents requires grounding, observability, resiliency, governance, and continuous monitoring. New challenges appear: Challenge Impact Hallucinated responses Reduced trust Missing citations Verification issues Tool failures Broken workflows High token consumption Increased cost Latency spikes Poor user experience Limited monitoring difficult troubleshooting Security concerns Compliance risks The lesson was clear: Production readiness is not about making the agent smarter. It is about making the agent reliable. My Observation While experimenting with AI agents in Microsoft Foundry, I found that model selection was rarely the primary challenge. Most production issues stemmed from grounding quality, tool reliability, observability, and security controls. Addressing these operational concerns often had a greater impact on user trust than changing the underlying model. Production AI Agent Reference Architecture A production-ready AI agent typically includes several components beyond the language model itself. Layer Technology User Interface Web App, Teams, Copilot Agent Runtime Microsoft Foundry Agent Service Knowledge Layer Azure AI Search Foundation Model GPT-4o Tool Integration APIs and Functions Monitoring Azure Monitor Security Managed Identity and Key Vault Request Flow Lesson 1: Start with the Use Case, Not the Model One of the most common mistakes is beginning with model selection. Many teams ask: Which model should I use? Should I choose GPT-4o? Should I use a larger context window? A more important question is: What business problem are we solving? Before evaluating models, define: Users Business goals Success metrics Compliance requirements Operational constraints An enterprise support agent and a financial compliance agent may use the same model but require completely different architectures. Key Takeaway Successful AI projects start with business outcomes, not model benchmarks. Lesson 2: Grounding Matters More Than Prompting Prompt engineering helps. Grounding drives trust. Without access to reliable enterprise data, even advanced models can generate confident but incorrect responses. Grounding Sources Azure AI Search SharePoint documents Internal knowledge bases Structured enterprise data Approved policy repositories Example from azure.ai.projects import AIProjectClient # Configure grounding with Azure AI Search agent = client.agents.create_agent( model="gpt-4o", name="support-agent", tools=[ { "type": "azure_ai_search", "azure_ai_search": { "index_name": "support-kb", "endpoint": "https://your-search.search.windows.net" } } ] ) In production environments, grounding through Azure AI Search helps agents retrieve trusted enterprise information rather than relying solely on model knowledge. This significantly improves response accuracy and trustworthiness. Try it yourself: Configure Azure AI Search as a knowledge source in Microsoft Foundry and compare grounded versus non-grounded responses. Key Takeaway Reliable retrieval is usually more valuable than sophisticated prompting. Lesson 3: Design Tool Usage Carefully AI agents become powerful when they can interact with external tools. Examples include: CRM systems Databases APIs Ticketing systems Business applications However, every tool increase complexity. Ask yourself: When should the agent call the tool? What happens if the tool is unavailable? How should failures be handled? Example Failure Scenario +-----------------------------+ | User asks for invoice status | +-------------+---------------+ | v +-----------------------------+ | Agent calls Finance API | +-------------+---------------+ | v +-----------------------------+ | API timeout detected | +-------------+---------------+ | v +-----------------------------+ | Fallback response returned | +-----------------------------+ Example Failure Scenario: A production-ready agent should gracefully handle external service failures and return a fallback response instead of failing completely. Key Takeaway Design for failure before designing for capability. Lesson 4: Evaluate Before Deployment Many teams test only for response quality. Production agents require broader evaluation. Evaluation Area Why It Matters Accuracy Correct answers Grounding Quality Faithful responses Latency User experience Safety Risk reduction Cost Sustainability Tool Success Rate Reliability Evaluation should become part of every deployment pipeline. Key Takeaway You cannot improve what you do not measure. Lesson 5: Make Observability a First-Class Feature Observability is often neglected until something breaks. Unfortunately, production systems always encounter unexpected behavior. Track metrics such as: Metric Purpose Request Volume Demand tracking Average Latency Performance Token Usage Cost visibility Grounding Success Rate Quality Tool Failure Rate Reliability User Satisfaction Business value Setting Up Tracing in Foundry from azure.ai.projects import AIProjectClient from azure.monitor.opentelemetry import configure_azure_monitor # Configure Azure Monitor for tracing configure_azure_monitor( connection_string="InstrumentationKey=xxx" ) # Run the agent response = agent.run( thread_id=thread.id, instructions="..." ) Tracing provides visibility into how an AI agent processes requests, invokes tools, and generates responses. By integrating Azure Monitor, teams can track latency, identify failed tool calls, analyze token usage, and troubleshoot unexpected agent behavior. This observability is essential for operating AI agents reliably in production environments. Try it yourself: Enabling tracing and monitoring for a Microsoft Foundry agent using Azure Monitor. Example Dashboard +-----------------------------+ | Production Monitoring Dashboard | +-----------------------------+ | Requests Today | 5,120 | | Average Latency | 3.2s | | Grounding Success | 97% | | Tool Failure Rate | 1% | | Average Tokens | 2,800 | +-----------------------------+ Monitoring production metrics helps teams understand agent performance, reliability, and cost efficiency. Key indicators such as latency, grounding success rate, tool failure rate, and token consumption provide valuable insights into the operational health of an AI agent and enable proactive troubleshooting before users are impacted. Key Takeaway What you can't observe, you can't effectively operate or improve. Tracing is a critical capability for maintaining production-ready AI agents. Lesson 6: Monitor Token Consumption Token usage directly impacts cost. A highly successful agent can quickly become expensive if token growth is unmanaged. Common optimization techniques: Optimization Benefit Prompt Compression Lower token usage Response Caching Reduced model calls RAG Filtering Focused context Context Trimming Smaller requests Model Selection Cost control Key Takeaway Cost optimization should be planned from day one. Lesson 7: Build Security into the Design Security should never be an afterthought. Enterprise AI systems must enforce the same security boundaries as traditional applications. Recommended Controls Control Purpose Managed Identity Secure authentication Azure Key Vault Secret management RBAC Authorization Audit Logs Compliance Content Filtering Safety Private Endpoints Network Security Guiding Principle An AI agent should never access data beyond a user's permissions. Key Takeaway Security is a design requirement, not a deployment task. Lesson 8: Expect Tool Failures External dependencies inevitably fail. Production-ready agents should anticipate: API downtime Authentication failures Network interruptions Rate limiting Unexpected responses Recommended Strategy Key Takeaway async def call_tool_with_resilience(tool_name, params): try: result = await tool_client.execute( tool_name, params, timeout=5.0 ) return result except TimeoutError: return await cache.get_fallback( tool_name, params ) Example Outcome: Introducing retry and fallback logic can significantly improve reliability. By handling temporary API failures gracefully and returning cached responses, when necessary, agents can reduce user-facing errors and provide a more consistent experience. Lesson 9: Evaluate the Process, Not Just the Output A correct answer does not always mean the process was correct. In production environments, teams should evaluate not only the final response but also the steps the agent took to generate that response. An answer may appear accurate even when it was based on incorrect retrieval results, unnecessary tool calls, or incomplete citations. Review: Retrieval quality Tool execution Citation accuracy Security compliance Reasoning path User Query | v Retrieval | v Tool Execution | v Reasoning | v Response For example, an agent might generate the correct answer by chance, even though it retrieved irrelevant documents or used an inefficient workflow. Without evaluating the process, these hidden issues may go unnoticed until they affect reliability, compliance, or user trust. Key Takeaway Inspect the entire workflow, not just the final answer. Lesson 10: Think Like a Production Engineer As adoption grows, operational excellence becomes the differentiator. Ask questions such as: Can we troubleshoot failures? Can we measure business impact? Can we control costs? Can we scale safely? Can we govern usage? These questions become more important than model selection over time. Key Takeaway Production success comes from engineering discipline, not model size. Microsoft Foundry Features That Helped Improve Reliability Foundry Capability Production Benefit Agent Service Agent orchestration Knowledge Sources Grounded responses Evaluations Quality measurement Tracing Workflow visibility Model Catalog Model flexibility Safety Systems Risk mitigation These capabilities help teams move beyond proofs of concept and build solutions that are ready for real-world adoption. Production Readiness Checklist Before releasing an AI agent, verify: ✅ Business goals defined ✅ Grounding strategy implemented ✅ Security controls enabled ✅ Evaluation pipeline established ✅ Monitoring configured ✅ Failure handling tested ✅ Cost optimization reviewed ✅ Governance process defined ✅ User feedback loop available ✅ Deployment rollback strategy prepared Get Started with Production-Ready Agents If you're building AI agents using Microsoft Foundry, start by focusing on grounding, observability, security, and evaluation from day one. Suggested next steps: Build your first agent in Microsoft Foundry Configure Azure AI Search grounding Enable tracing and monitoring Evaluate agent quality before deployment Add resilience and fallback strategies Which of these 10 lessons has been most valuable in your own AI agent journey? Share your experiences and insights in the comments.Model router updates: new regions, a refreshed model pool, and understanding the hill climb
Across Microsoft, "hill climbing" has become shorthand for how real AI progress happens: not in one dramatic leap, but through a disciplined loop. Microsoft AI defines the hill climb as an organization that continuously improves, cycle after cycle, through more compute, better data, and sharper evaluation. Reinforcement fine-tuning in Foundry defines it as improving the deployable model package one measured step at a time across quality, latency, and cost. Different altitudes, same premise: progress is not a one-shot decision. It's a loop. For most teams, the decision of what model to use when is made manually or with custom routing tools. A developer picks a model based on benchmarks, familiarity, or the last launch that made headlines, ships it, and revisits the choice only when something breaks. In an ecosystem where the frontier moves monthly, that decision goes stale fast. Model router in Foundry Models brings the hill climb to the selection layer. What's new: a bigger pool, in more places This release expands where teams can deploy model router, broaden the supported model pool, and delivers updates through a stable endpoint. Together, these changes help teams run production workloads in more locations, match a wider range of tasks to suitable models, and adopt supported updates without changing the application integration. A refreshed model pool. The supported model list now includes Anthropic Claude Opus 4.8 — a high-capability model built for complex reasoning and long-form generation, for scenarios that demand depth, structure, and quality — and the GPT-5.6 family. Just as importantly, the pool is pruned: gpt-5-chat, gpt-5.2-chat, gpt-5.3-chat, Deepseek-V3.1 have been removed from the model router as models reach the end of their lifecycle and are deprecated in Foundry. New region availability. The model router is now available in 28 regions for global standard and 21 data zone regions. For many organizations, inference requests must stay within specific geographic boundaries for regulatory, governance, or customer-trust reasons — and intelligent routing shouldn't force a compromise on that. Find the full list of regions here. The most important detail is what you don't have to do: these updates occur automatically*. The endpoint remains stable as the supported model pool is refreshed, so teams do not need to redeploy the model router to receive the update. Applications can continue using the same integration while the model router evaluates requests against the current supported pool. Teams should continue monitoring routing traces and application outcomes to confirm that quality, cost, latency, and governance requirements are met. *Models from Anthropic still need to be deployed separately before they can be routed to through the model router. Interested in hearing more about what's new to the model router? Tune in for the next episode of Model Mondays with Sanjeev Jagtap and Lee Stott, where they talk all things model router from evaluations to hill climbing. Sign up here to watch live or view the replay: Model Mondays - Spotlight On Model router in Microsoft Foundry | Microsoft Reactor The selection-layer hill climb At the selection layer, a step is a routing decision. Each one is a micro-optimization against your objective, and each one is instrumented: every response from the model router includes a model field showing which underlying model was selected, so the climb leaves a complete, auditable trail. Model router supports three parts of the optimization loop: A/B testing to compare two router configurations to understand quality, cost, and latency tradeoffs; model decomposition to use routing results to decompose a single-model application into a multi-model or multi-agent design, and continuous routing to keep the router in production for continuous per-request selection. Each pattern turns model choice into a measured, repeatable process rather than a fixed decision. 1. A/B Testing Question: Which model or routing strategy should I use in production? A/B testing helps teams compare candidate models, model families, or router configurations against the same workload. Representative traffic is sent to competing deployments, and teams compare quality, cost, latency, and governance outcomes. The goal is to understand tradeoffs and identify the model or routing strategy that best meets workload requirements before promoting it to production. 2. Model Decomposition Question: What work is my application actually doing? Model decomposition uses model router as a diagnostic tool. By deploying the model router against a representative workload and examining routing telemetry, teams can see how requests naturally separate into different task classes. Simple retrieval, classification, and summarization requests may route to smaller models, while reasoning, planning, and agentic workflows may require more capable models. The goal is not to choose a winner, but to understand the structure of the workload and uncover opportunities for optimization, specialization, or architectural improvements. 3. Route continuously Question: Why choose a single model at all? Route continuously is the pattern model router was designed for but is not limited to. Rather than treating model selection as a one-time decision, teams leave the model router in production and allow the best-fit model to be selected for each request. As the supported model pool, regional availability, and platform capabilities evolve, teams can continue using the same endpoint while evaluating whether updates improve workload outcomes. Model selection becomes an ongoing optimization process rather than a project that must be repeated every time the model landscape changes. Together, these patterns illustrate a broader shift: the model router is more than a model. It is a tool for the optimization loop itself, helping teams evaluate tradeoffs, understand workload behavior, test hypotheses, and continuously refine model selection as requirements evolve. Whether used to compare candidate models, decompose applications into specialized tasks, or automate per-request routing in production, model router turns model selection into an observable, measurable, and repeatable process. As the model landscape continues to change, that optimization loop becomes a durable advantage. Getting Started Ready to start your own hill climb? Whether you're exploring the model router for the first time, evaluating routing strategies against your workload, or building a long-term optimization practice, these resources can help you move from experimentation to production with Microsoft Foundry. What's new in model router? Sign up for the next Model Mondays episode for a deep dive into new features, optimization patterns, and the latest model router updates. How do I build agents with model router? Check out the Model Router Agents Lab and build agent experiences with routing, retrieval, web search, tool calling, and multi-agent patterns. How do I evaluate model router? Compare model router against baseline models using your own prompts, then review quality, cost, latency, and routing decisions with the Auto Evaluation Toolkit. How do I optimize model router for my workload? Start your hill-climbing journey with the Model Mastery workshop, where you'll test one optimization lever at a time and measure how each change impacts workload outcomes. How do I build a model router optimization playbook? Explore the Model Releases repository to track new capabilities, understand the optimization question behind each release, and try focused notebooks that demonstrate one optimization lever at a time.
