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, Microsoft 365 Copilot, WorkIQ, Foundry, and Fabric, Azure Copilot, are addressing this need with standardized agent observability data built on OpenTelemetry and its GenAI semantic conventions, with availability outlined below. Unified agent observability in Azure Monitor connects that data across platforms so developers and site reliability engineers can follow execution, investigate cost, evaluate responses, and debug failures.
An employee asks an agent,
- Which customer orders are at risk?
- Which tool or knowledge source did the agent use?
- Where did the turn fail, or spend time?
- Did the agent delegate work to another agent?
- What model and tool activity contributed to the result or cost explosion?
Answering may involve several agents, enterprise data, and external services. When the response causes token explosion or is slow or incomplete, developers and site reliability engineers need to understand the entire journey.
Unified agent observability in Azure Monitor Application Insights Agents (preview) connects execution across agent platforms and summarizes agent runs, tools, and models, helping teams follow a request, locate failures, and investigate dependencies without stitching logs together by hand. Agent owner can also understand the agent reasoning logic and correlates issues across agent platforms with the e2e traces which can be viewed agent graph, Azure observability skills.
Trace across the agent ecosystem
|
Agent or platform |
Tracing support |
|
Copilot Studio, Foundry, and GitHub Copilot agents |
Supported Setup: Copilot Studio Environment-level telemetry with Application Insights (preview)
|
|
External agents following OpenTelemetry GenAI semantic conventions |
Supported with compatible telemetry export and trace context propagation Setup: Register external agents in Foundry, External agents in Foundry (GA), |
|
Fabric agents |
Coming in December 2026 |
|
Microsoft 365 Copilot agents and WorkIQ |
Supported via the A365 exporter, currently in preview Set up: Sync agent activity from Agent 365 to Azure Monitor (preview), |
With telemetry collection enabled and trace context preserved across handoffs, teams can investigate connected activity in Azure Monitor Application Insights. Each trace reflects the participating components and the telemetry they emit.
From an incomplete answer to the failing dependency
Consider an illustrative workflow: a Copilot Studio agent delegates a task such as order analysis to a Foundry agent, which queries an instrumented inventory service and calls an external agent to check delivery options.
If the recommendation is incomplete, an engineer can follow the request in Azure Monitor Application Insights. The trace might reveal that the delivery lookup timed out, showing where the failure occurred and how long it took. Where additional detail is emitted and collection is enabled, tool inputs and outputs or model interaction metadata provide further context.
This connected view helps teams identify which agents and tools participated, where time was spent, and which operation needs attention.
See Demo
Built on OpenTelemetry
OpenTelemetry trace context connects operations across components, while GenAI semantic conventions provide common descriptions of agent and model activity. External agents can participate through compatible instrumentation and telemetry export, extending visibility across Microsoft and external runtimes.
Start with a customer workflow that crosses agent boundaries. Enable telemetry, verify correlation across handoffs, and exercise both successful and failing requests. The goal is practical: give the team operating the application enough evidence to understand what happened and act.
Technical references