michelangelo

michelangelo is a research agent harness — a workspace where many AI agents work the way a research group does: running across your machines, set on a problem, and checking one another until something holds.

What does the future of science research look like?

Software spent the last few years building harnesses for its agents. Science mostly stopped at the chatbot. michelangelo is the harness for what comes next.

Goal Agents fan out Reason → claim Reproduce → evidence + − Δ ε ε≈0? yes Knowledge no
A goal enters. Agents fan out — some to reason toward a claim, some to reproduce the evidence — and the two are checked against each other. Only what survives that check becomes knowledge; the rest loops back to be refined or refuted.

Underneath is a simple commitment: you own the system, and the system is text. michelangelo lives in the tradition where your whole environment is plain configuration you write and reshape — the “everything is a config file” lineage of Arch Linux, and lately of DHH’s omarchy. A config is not a settings menu; it is the system itself, written down — so it bends to how you actually work, and runs the same anywhere.

On that footing, agents stop being a single chat box. They become workers you can watch, talk to, and set against one another; every machine you own behaves as one computer; and the way your group likes to work is itself just configuration. Built with Claude, and built for Claude to drive.

Agnostic to any one research direction, and grounded in our own group’s work — where verifying scientific claims is itself a line of research.