Start with the questions a review must answer

A tool-using agent receives a task, assembles context, selects tools and may ask a person to approve an action. The final chat message usually compresses that work. It can omit denied calls, retries, intermediate source reads, edited proposals and an external action that completed after the model stopped responding.

After an unexpected result, a reviewer will need answers to questions such as:

• Who started the task and under which identity did it run?
• What objective and limits were in force?
• Which private objects did the agent read?
• Which model, instructions and tool definitions shaped the run?
• What action did the agent propose, and what exactly was approved?
• Which calls reached an external system?
• What did each destination confirm?
• Was anything left in an uncertain or incomplete state?
• Can the reviewer reconstruct the sequence without opening sensitive content unnecessarily?

Answer these questions from structured events. Keep the transcript as one source beside tool events, policy decisions, approvals and destination receipts.

Decide the review audiences as well. An operator recovering a failed task needs a short sequence and current status. A security reviewer may need denied access attempts and identity details. A user may need a plain account of records viewed and changes made. They can use the same event stream through different views and access rules.