Human · Task prompt
Task given
The human asks for a fix: formatRelative calls a note written at 23:59 “today” when it is read at 00:00. The test stays as it is.
AI agent session replay
Flight Recorder
Scrub through a coding agent’s session like a black box: what the model decided, what the harness enforced, and where a human stepped in.
Fictional session · written for this page, not a real log
Choose a session
The midnight test
Task: A date test fails only when the suite runs just after midnight. Fix the code, not the test.
Where time and tokens went
Longest step: Review gate: PASS at 00:28 (31 s). Most tokens: Tests pass at 00:21 (4,906).
Drag the timeline or use the arrow keys, Home and End. Playback pauses in a background tab.
Steps
Flight data
- Session time
- 59.2 s
- Model tokens
- 19,335 in · 1,238 out
- Tool calls that ran
- 3
- Policy stops
- 0 denied · 0 asked
- Gate verdicts
- PASS
- Changed
- 1 file · +3 −1
Steps by actor
- Model5 steps · 28 s
- Harness2 steps · 31.2 s
- Human1 step · 0 ms
What to notice: Even a clean run leaves evidence: the failing test that proved the bug, the diff, and a gate verdict nobody had to take on trust.
How to read the replay
- Model chose a plan, a tool call or an edit.
- Harness enforced a policy rule or a review gate.
- Human gave the task or answered a question.
Why a harness at all? OWASP lists excessive agency (too much functionality, permission or autonomy) as a top risk for LLM applications, and recommends least privilege plus human approval for high-impact actions.
Why this room
Umut Korkmaz builds agent tooling: Quorate, an open-source CLI that runs several AI reviewers and ends in one PASS, WARN or FAIL verdict, and Convoke, a Rust terminal coding agent with Quorate review gates. This room shows the kind of evidence trail those gates are built to leave.
See both in ProjectsReplay boundaries
All three sessions are fictional and were written for this page. The repository, files, commands, package names, timings and token counts are illustrative; no model was called and no real log was used. The policy rules use a made-up syntax. Gate verdicts follow Quorate’s rule: FAIL when a finding reaches the fail-on severity (high by default in its GitHub Action), WARN when only lesser findings remain, PASS with none. Quorate also reports WARN for a degraded run, one in which no real reviewer finished; none of these runs is degraded. In this replay a WARN also waits for a person.