Changing a prompt can produce a better answer without fixing the problem. Agents can behave differently from one run to the next, so one good result is a poor test of a fix.

We’ll discuss how to use an agent’s recorded inputs, tool calls, and results to investigate failures, then turn those findings into repeatable evaluation cases. Those cases define the behavior you expect, so changes can be checked against known failures and cases that already work. We’ll compare latency and cost alongside results from repeated runs.

That evidence helps decide whether to change instructions or retrieval, fix a tool, or limit what the agent can do without approval. Each investigation should leave you with a regression check you can rerun when the agent changes.