An instrument, not an argument

You install a coding agent. Then a notes agent. Then something for your calendar. They share your filesystem, your clipboard, your OAuth scopes, your MCP servers, your wallet. Nobody has systematically studied what happens when they interact.

CRUCIBLE clones real agents from GitHub, pins their commit SHAs, builds them into containers, checks whether those containers start, and publishes the distribution of why the rest do not. Later phases run working pairs against each other under adversarial scenarios with deterministic criteria.

An estimate of a risk is an opinion about it. This runs the experiment.

Built on SMADP, not forked from it

SMADP published the estimated matrix of multi-agent risk and the sandbox that can execute it. CRUCIBLE imports that sandbox as a library and contributes the half that was missing: getting real agents to the starting line in volume.

Source

The engine, this site, every spec and every run record live in one repository. The specs include the arguments that were rejected and why, which is usually the more useful half.