It records everything the agent did, then rewinds to find the exact moment it first went wrong.
AI agents now work on their own for hours, making dozens of moves in a row. When one fails, you are left scrolling through thousands of log lines trying to guess where it slipped. There is no standard way to record what an agent was thinking, and no quick way to point at the one decision that doomed the whole run. It is like investigating a plane crash with no black box on board.
This is the black box. It clips onto any agent and quietly records every action, every tool it reached for, every observation, and every change in what it believed to be true. When a run fails, it replays that recording and pinpoints the earliest moment things went off track. Then it lets you ask plain questions about the run, like where did it first misunderstand the goal.
| Item | Description | Signal |
|---|---|---|
| Misread the goal | It got the task wrong at step one, and everything after inherited that mistake | Silent |
| Picked the wrong tool | It chose a reasonable looking but wrong tool, and the result seemed fine | Delayed |
| Passed along bad input | It handed off a broken value that only caused trouble 30 steps later | Delayed |
| Beliefs drifted | What it thought was true started clashing with something it had already seen | Building |
| Stuck in a loop | It kept retrying the same failing move without rethinking | Obvious |
| Forgot a key fact | An important detail slipped out of memory and was never looked up again | Building |
A set of five research projects on making AI agents reliable, understandable, and safe.