The Organism

The question

What happens if the agents can organize the research environment themselves, while the things that matter are governed by code?

Steve Yegge’s Fences, not Sandboxes gave me a starting point. I was already working with agents on Lean research. I wanted to take that further: less prescribing every step, more freedom to investigate, with a firm verification boundary.

What I built

The Organism is an agent-led research environment built around Lean verification. Agents propose and pursue research work; proof checks, statement reviews, and protected verification machinery govern what gets accepted.

Around that core, the agents built their own working environment: research contracts, mailboxes, records of mistakes, a task board, and visual artifacts for understanding the lab. The public essay documents this setup and its integration work with CRRG.

My contribution is the research direction and the constraints that shape the system, developed through repeated experiments and feedback with the agents. Much of the implementation is agent-built. That division of work is part of the experiment.

What changed my mind

The interface matters. Asking for a document explaining the lab led to a visual overview, then to a set of artifacts that became the main way to review and discuss its work. Keeping a human oriented in an evolving research system is a design problem of its own.

The other lesson is about verification: a kernel can check a proof, but a correct proof of the wrong statement is still the wrong result. Preserving the meaning of the question matters as much as checking the answer.

Still under construction

How do these systems compose when another human joins? What should be shared between personal research environments? How much useful autonomy can strong verification support?

These are live questions. The public write-up is a snapshot of the experiment, rather than a claim that autonomous research is solved.

Read the full field report →