Inspect the individual tool steps
Open the run's recorded steps and locate each tool action. Check which tool was invoked, what input it received, the summary of its activity, and whether an error was recorded. When available, inspect the result metadata too, such as the page or resource returned. This gives you a concrete sequence to compare with the task you assigned.
Do not treat a short summary as a complete audit of every side effect. A tool may report that a request was sent or a file was written without establishing that the recipient processed it or that the file contains the intended content. Verify important outcomes at the destination, especially when the action affects a customer, a live system, or shared data.
Use the history to decide the next step
If a step failed, compare its input with the error and check whether earlier steps already produced changes. Correct the specific issue before asking the agent to repeat work. If the step succeeded but the final answer omits a requested detail, inspect the recorded result and ask for that missing verification explicitly.
For recurring workflows with agents, retain enough run context to understand what happened without exposing unnecessary sensitive data. A clear record of inputs, tool actions, and outcomes makes troubleshooting more reliable and helps a reviewer distinguish the agent's interpretation from the tool's actual response.