AGENT THOUGHT
ACTION
Browse the vocabulary, then follow the evidence in real agent trajectories.
Start with an action. Follow it down to a concrete behavior.
Explore two SWE-agent tasks and a SWE-chat session shaped by human feedback.
Look beyond the final answer. In the requests trace, the agent runs a reproduction script and inspects the result. In the django trace, it submits after acknowledging that it has not verified the change.
Follow the human intervention. After a successful build, H3 questions why the table definition is duplicated. At T22 the agent reconsiders; at T23 it reverts the baseline change.