Read an agent session.
Understand what agents did, how they delegated work, and why a session was flagged. Follow this Beams example from the summary to the recorded evidence.
Read the summary.
Start with What happened for an overview of the task and its reported outcome. Coverage shows how many recorded calls were analyzed; classifier matches show what needs a closer look.
This example spans three conversations, eight agents, and 222 calls. The agents fixed a deployment script, changed a web server configuration, and corrected a unit test.
Follow the timeline.
Read messages and tool activity in order. Expand a tool group to inspect its calls. Branches show where an agent delegated work to another agent.
Here, the main agent runs deploy.sh, then starts an independent review
with Arendt. The branch connects the main agent's tool activity to the delegated
review.
Inspect the evidence.
Open a classifier finding to see the matched behavior, its risk level, and the recorded calls cited as evidence.
The secrets-exposure finding points to tool call #13: a deployment script
printing an example AWS access key ID. The evidence gives the reviewer something
concrete to verify.
Classifiers evaluate a session after summarization. A match calls for review; it does not prove intent or block an action.
View the full session screenshot
