How the lab works: AI agents as the researchers
I have more research ideas than time to run them. So I set up a lab where the agents do the experiments and I direct. I talk an idea over with a model until it's sharp, hand it to a long-running agent, and let it read the literature, write the code, run the experiment, and write up whatever survives. Most ideas don't survive. That's the point: one person can keep a dozen threads going at once, cheaply, and find out fast which ones were wrong.
How a project runs
I have an idea and discuss it with a model. It pokes holes, digs up the closest prior work, and we turn it into something you can actually test. Then I point a long-running Claude Code or Codex session at it and let it go. When it hits a real fork — a strategic call, an expensive job, a result that looks too good — it stops and asks me. I decide, and it keeps going. Some projects I preregister carefully; others start as a half-formed hunch I throw at an agent on a Sunday. Either way, while I'm asleep several agents are each a few experiments deep on ideas I'd never have had time to touch.
Everything lives in files
The lab lives in markdown, not in any model's head. Every project folder keeps a plain record: the aim, the log, the results with their exact configs, the open questions. Agents are good at producing work and bad at remembering it across sessions, so an experiment that ran but never got transcribed, with its numbers, didn't happen as far as the next agent is concerned. And the next agent is often a different model, days later, picking up where the last one left off. The files are the only memory that survives that handoff.
Ideas get killed
For some projects I write down what would count as failure before I can see the result, then honor it. One pilot's gate required a calibration rejection rate inside [0.045, 0.055]; a fresh run came in just under the line, and I dropped the project before it produced a single claim. A continual-learning idea died when the term it hinged on turned out to add no predictive power. A video world-model idea failed on paper — the arithmetic said the speedup couldn't exist — before any training ran. Those kills are why I trust the six write-ups. A lab where nothing dies has quietly loosened its standards to keep things alive.
What this is and isn't
This isn't agents doing research unsupervised, and it isn't a product. I supply the ideas, the direction, and every commit to git; the agents supply the breadth — reading, coding, running, and attacking far more than I could alone. It's small and compute-poor on purpose, which makes it better at auditing and falsifying than at building anything big. That's why so many of the results are careful negatives. If a number in one of these articles looks too clean, it's because it had to survive being written down and argued with first.