AI agents are becoming easier to build. Understanding why an agent gave an incomplete answer, where it spent time, or what drove its cost is harder, especially when requests cross agents, tools, and runtimes you didn't build.
For managed agents, you can't instrument the platform's internal code yourself. Runtime providers must expose compatible telemetry, and that telemetry needs to connect across platforms.
At Microsoft Ignite 2026, Azure Monitor and Microsoft Foundry expand agent observability around one improvement loop: observe execution and investigate failures, evaluate response quality, then optimize and validate changes. Microsoft Foundry helps agent builder build, evaluate, and improve agent quality, while Azure Monitor, via Application Insights, provides IT operators deeper visibility into agent performance, reliability, and cost full stack across applications, infrastructure, and agents built on any platform. OpenTelemetry connects agent diagnostics with application and infrastructure signals, giving builders and operators shared evidence.
Highlights include cross-agent debugging in public preview, external-agent observability and evaluation in Foundry & Azure Monitor becoming generally available, Azure Monitor traces and Trace Analytics entering public preview, and Azure Monitor Health Models becoming generally available. Coding-agent workflows bring investigation into development; content access controls help protect sensitive telemetry.
Connect agents across runtimes with OpenTelemetry for builders and operators
Cross-agent debugging, in public preview
Creating an AI agent can take minutes, with little or no code. Understanding its behavior, tracking its cost, and debugging a response across multiple agents is much harder—and sometimes impossible without visibility into the runtimes behind them.
Your agent may rely on managed runtimes and components for which you have never written a line of code. You cannot simply add instrumentation inside them. You depend on their providers to expose the telemetry needed to understand what happened.
Microsoft agent runtime and framework providers—including Copilot Studio, Foundry, GitHub Copilot, Azure Copilot, WorkIQ, M365 Copilot, CoWork —are addressing this need with standardized agent observability data built on OpenTelemetry and its GenAI semantic conventions.
Azure Monitor and Microsoft Foundry share one platform and the Microsoft OpenTelemetry Distro. Unified agent observability in Azure Monitor connects that data across agent runtime and frameworks, so developers and SRE can follow execution, investigate cost, evaluate responses, and debug failures.
For agents outside Foundry, external-agent observability and evaluation becomes generally available. This extends to agents built with other frameworks and platforms and hosted anywhere, including on AWS and Google Cloud (GCP). Teams can centrally observe, evaluate, and optimize these agents, with their traces centrally managed in Azure Monitor. Builders instrument agents to send compatible traces to Application Insights and register them with Foundry, keeping existing endpoints. Registration, agent trace view, Agent Insights, and continuous evaluation of production traces bring them into the same workflow. Registration isn't instrumentation and doesn't grant source-code access.
Learn more: Cross Agent Obs Demo, Unified agent observability with end-to-end tracing in Azure Monitor, Copilot Studio environment-level telemetry, Register external agents in Foundry, External agents in Foundry (GA), Sync agent activity from Agent 365 to Azure Monitor (preview), Agent 365 Blog.
Observe and evaluate: understand behavior, quality, and cost
Once telemetry is connected, teams need to understand how an agent runs and how well it serves the customer. A successful tool call doesn't necessarily produce a useful answer.
Observe behavior. Agent details in Application Insights and Foundry show per-agent metrics, errors, tool calls, models, and tokens, correlated with application and infrastructure signals.
For agents working with more than text, inspection of images, PDFs, and audio in agent interactions is in private preview, extending the evidence available to understand a run.
Evaluate quality. A new evaluations view brings Foundry evaluations into Azure Monitor Application Insights, so operators can track quality alongside production behavior. In both Application Insights and Foundry, annotations capture human feedback on runs, and a safety evaluator scores responses against safety rubrics. Scores need interpretation against each evaluator's rubric, not just comparison on a normalized display.
Learn more: Monitor AI agents in Application Insights and Run evaluations in Foundry.
Observe across the stack: investigate interactions and workload health
A poor response or slowdown may originate in a model interaction, a tool, an application dependency, or infrastructure. The observe stage extends across that execution path.
Azure Monitor Tracing Platform and Trace Analytics, in public preview, build on Application Insights distributed tracing to accelerate investigation at scale. Azure Monitor tracing platform assembles and stitches spans across resources into key transactions and compares healthy and failing transactions, including GenAI interactions.
Critical Path Analysis separates active work from waiting, allowing spot patterns and insights quickly out of thousands of traces. Attribute Analysis highlights differences between selected traces and a baseline. Trace-derived metrics connect transaction volume, duration, and errors to service-level indicators (SLIs), alerts, and dashboards.
Azure Monitor Health Models, now generally available, connects resource dependencies and operational signals to team-defined workload health criteria. Tracing explains what happened and where; Health Models helps operators assess customer impact and prioritize action across supporting services.
The traces preview requires onboarding and has regional restrictions. Review current prerequisites and limitations.
Learn more: Azure Monitor traces and Trace Analytics and Azure Monitor Health Models overview.
Optimize: improve agents in your existing workflow
Observability becomes useful when teams act on what they learn and measure whether the change helped.
Foundry supports skill optimization from evaluation results, suggested fixes for failing runs, and cost optimization to reduce token consumption and model spend. Suggested code changes for agents are coming soon.
The Azure Observability Skills Plugin, available through Azure Skills Marketplace and Awesome Copilot, brings Detect, Investigate, Remediate, and Validate into coding-agent workflows. It helps discover capabilities, confirm access, and retain investigation evidence. Azure Copilot's Observability Agent also supports troubleshooting, including high token usage.
Azure Monitor dashboards with Grafana MCP supplies telemetry access through a remote Model Context Protocol (MCP) endpoint. Agents can query Azure Monitor metrics and logs, Azure Resource Graph, Prometheus, Azure Data Explorer, and Health Models, and find, read, or update dashboards. Calls use the signed-in Microsoft Entra ID user's permissions; connecting doesn't grant additional access. Operational changes remain subject to applicable approval.
Learn more: Agent Insights in Foundry and Azure Monitor dashboards with Grafana MCP.
Fabric Agents and Apps observability, in public preview, applies this development workflow to Fabric agents and applications with built-in platform health and application telemetry. Developers can use their coding agent to enable frontend telemetry and investigate usage or errors in Application Insights alongside application code. This preview is separate from upcoming Fabric Data Agent tracing; Monitoring Hub extensions shouldn't be assumed available initially.
Related announcement: Observability in your coding agent for Fabric Apps.
Control access to sensitive agent content
Agent telemetry can contain prompts, responses, and sensitive information. Teams need control over collection and access throughout the improvement loop.
Protected tables, in public preview, provide separate role-based access control (RBAC) for GenAI content such as prompts and responses. Teams configure content routing together with table protection; protecting new content doesn't protect historical copies elsewhere.
Personally identifiable information (PII) redaction, in public preview, detects and masks configured patterns such as email addresses and Social Security numbers in Log Analytics. It complements access controls and content-capture policies, but doesn't guarantee removal of all sensitive content.
Learn more: Protect GenAI content and PII redaction in Azure Monitor logs.
Next steps:
Choose a customer workflow to improve. Verify runtime support and content-capture settings, connect telemetry, follow an execution, and evaluate response quality. Make a targeted change and validate its effect on quality, reliability, latency, and cost.
Start with Azure Monitor Application Insights for full stack agent production visibility, Foundry for evaluation and optimization, or Trace Analytics for investigation across services. Review availability and preview limitations before adoption.
A lower-cost SKU for agent telemetry, with a starter kit and free tier, is coming soon.