Rootstep

Find the step where your AI agent's failure began.

An AI agent tried to fix a bug and failed. Rootstep went back to the save point before each step, let the agent decide again several times, and measured how often the run could still be saved. Where that number falls to zero and stays there is the point of no return.

Loading…

How it works

1. Record

Every command the agent runs happens in a Nebius Token Factory Sandbox, which saves a snapshot of the files after each step.

2. Go back

For a step k, Rootstep loads the snapshot from just before it and gives the agent the same conversation. NVIDIA Nemotron decides step k again, several times in parallel.

3. Measure

The share of re-runs that succeed is the step's rescue rate, shown with a 95% range. The last step that can still be saved marks the point of no return.