When AI agents make mistakes, developers that created them need to manually reconstruct failures and search for patterns in the “traces” or logs of activity that capture the work involved in completing a task.
Microsoft is striving to simplify that manual process with a new feature in its Foundry AI development and governance platform. Insights in Foundry analyzes traces to determine likely causes of agentic mistakes and recommend ways to overcome them. It surfaces recurring and previously unknown behaviors, explains likely causes with supporting evidence, and recommends next steps for developers to resolve them.
Foundry already provided tracing, monitoring, and evaluation capabilities to help development teams understand where agents need attention, evaluation, or more extensive iteration.
Insights in Foundry goes deeper, analyzing volumes of traces (which are stored in the Azure Monitor Application Insights resource) and looking for patterns that would be difficult to find by reading one trace at a time.
Patterns yielded by Insights give developers a starting point for investigation. Findings can include failures in tool calls, context handling issues, output quality, reliability, latency, or unnecessary token use.
Possible responses to a confirmed finding include:
- Making a targeted change to correct an instruction, tool description, code path, or workflow when evidence indicates a specific problem
- Adding evaluation coverage to create or update a dataset so the behavior becomes an explicit test before and after a change
- Routing the issue to its owner if a dependency or platform failure is the cause, then working with that owner to fix the underlying issue
- Exploring alternatives by using an optimization workflow such as Agent Optimizer — which generates improvements to evaluate then measures them against developers’ goals — when improving the behavior involves generating and comparing multiple candidate solutions.
Insights doesn’t open pull requests, build evaluators, or deploy fixes on its own. No updates occur without human review. Developers review each proposal and then manage through their normal processes including source control, evaluation, approval, and deployment.
To illustrate how Insights will operate, Microsoft provided the example of an AI agent that performs customer order support. If an order-status lookup fails, an agent that’s not functioning properly may still tell a customer that their order was delivered. A single manual trace could make that look like an isolated mistake. Across hundreds of requests, it becomes a pattern requiring deeper evaluation.
Insights will analyze order support agent traces, which show the lookup failure followed by the unsupported delivery claim. A developer can inspect the request, the tool calls and their results, as well as the final response, then compare them with successful executions. The development team might then test whether an instruction change prevents unsupported delivery claims while preserving correct answers when the lookup succeeds. They validate the result before deploying the change and review new production evidence afterward.
Insights in Foundry is available in public preview.
More Insights on AI Agent Development:
- Microsoft, GitHub, and Anthropic Reshape AI Developer Economics
- Developer Economics: GitHub Delivers Multi-Model Orchestration at Runtime
- With GitHub Copilot Integration, Teams Powers Greater Developer-Agent Collaboration
- Microsoft Governance Tools Ensure AI Agents Play Within the Rules at Runtime
