Agent activity

Agent activity watches the coding agent's own tool calls and sessions, and reports the patterns that mean it is in trouble. Loops read it to notice an agent that is stuck, or that changes code without evidence, before its output reaches a review.

What it sees

It has two sources. Hooks see every tool call as it happens, early enough for a loop to refuse the call. Session transcripts are read afterwards and show trends across many sessions. It sees what the agent did, not what the code does: an agent that works in good order can still produce wrong code.

Reading the signal

• No pattern: nothing in the calls or sessions read matches a known sign of trouble.
• A named pattern: the pattern and its evidence, such as the same call repeated, an edit made with no failing test on record, a tool that keeps failing, or a run that stalled.

A gap in the record, such as a hook that did not run or a transcript cut short, is not a quiet session. Read missing data as missing, not as a pass.