Systems worth breaking.

We build small, useful AI systems, then stress them until the interesting parts fall out.

active experiments

public projects

∞

questions left

The lab loop

Useful things come from honest failure.

Start with a sharp question and a small surface area.

Push the system past its happy path on purpose.

Keep what works. Share what failed. Repeat.

Selected work

Composable primitives for deterministic agentic workflows.

Continuous security analysis for large, moving codebases.

A controlled arena for testing cooperation, conflict, and failure.

Research in public

No black boxes

The best tools get better when more people can inspect them, challenge them, and make them their own.