Agent replays

Watch a research agent work through a real task, one step at a time, and see how it was graded.

Recorded runs

Research reproduction audit
Pass
Research-agent run, recorded, de-identified
Actions
5
Held-out checks
6/6
Score
1.0
```
verdict: PASS
held-out checks: 6 / 6
score: 1.0
```
A captured research-agent run that builds and tests an auditor for a physics-informed model reproduction packet.
Watch it step by step See each tool call
Training diagnostic
Fail
Research-agent run, recorded, de-identified
Actions
24
Accuracy
0.785
Threshold
0.95
01.0
```
verdict: FAIL
submission format: PASS
held-out accuracy: 0.785
pass threshold: 0.95
```
A captured classification run that explores features, tests models and produces a valid submission, then falls short of the held-out accuracy threshold.
Watch it step by step See each tool call

About the replays

Each replay is one real attempt at a task. You can watch it step by step, or look at every tool call with what went in and what came back. We removed identifying details and some raw data before publishing.

The runs come from ML Research Tasks.

Request a task sample Agent environments