From traces to better agents: The Foundry improvement loop | BRK249
Takashi Okamoto, Yuji Shono, Charles Kim, and Avinav Jami present a production-focused approach to improving agents in Microsoft Foundry, moving from observed behavior to tested changes.
The agent improvement loop
Building an agent is only the starting point. This session focuses on the ongoing work of improving it in production by connecting traces, evaluations, and optimization in a repeatable process:
- Find high-impact issues: Use production traces to identify agent behavior that needs attention.
- Test improvements: Evaluate changes to determine whether they address the identified issues.
- Re-measure results: Measure behavior again after changes to assess their impact and guide further improvements.
The process addresses both agent quality and cost management, while keeping developers in control of the changes they test and ship.
Session details and resources
- Event: Microsoft Ignite 2026
- Session: BRK249 — Agent Platform
- Format: Breakout
- Level: 200 — Intermediate
For further learning, see the Agent Platform next steps resources.