Framework Guide

Put OpenClaw agents through a real pre-deployment trial.

If your agent already runs in OpenClaw, Crucible gives you the missing proof layer: survival scoring, replayable failures, and deployment-grade reports.

How this integration works

Wrap a turn function instead of rebuilding your agent loop.

Measure whether agents ask for help before guessing through ambiguity.

Use reports and hidden evals to decide whether an agent deserves more autonomy.

Starter example

from crucible import evaluate, wrap_openclaw_agent

wrapped = wrap_openclaw_agent(run_openclaw_turn)
result = evaluate(
    wrapped,
    name="Clawbot-EngineeringMode",
    scenario="capital-allocator",
    seed=11,
)