Probe-and-Resume critic gives agents 8.4 points more success
The critic does a check only when the agent tells it that the task is done.
In Probe-and-Resume, a critic does a check only when an agent tells it that the task is done. On Terminal-Bench 2.1, success was 8.4 percentage points more than MCTS with standard critics, at less than 50% of the cost of the Codex CLI agent.
Sources
Posted
