Long-running agents
I use my plasma physics research as a test bed for AI agents that work on hard problems for weeks. Physics suits the test because many of its answers can be checked. The agent's rules, its checks and its mistakes are all in a public repo, krmhd-research.
The posts below are grouped by what they teach. For the physics itself, see the research page.
Trusting work you didn't do
- Building a Gyrokinetics Code Without Reading a Single Line: rebuilding the simulation in JAX with Claude in 30 days
- Physics-Oracle Validation: how to trust code you've never read
- Authority to surprise us: the results I now trust came from tests that could have failed
How much to hand over
- Testing the Intelligence Explosion: the start of the experiment, asking whether AI can turn one physicist into a research team
- The Autonomy Gradient: what AI can and can't do across research tasks
- AI as the Undergrad Researcher: how the spring study ran, with AI doing the undergraduate-level research work. Written before the study was closed.
- The Boundaries of Delegation: what I hand over to AI in research, one year on
Writing and thinking with an agent
- Writing a Physics Paper with Claude: workflow, iterations, and hallucinations caught in review
- A Week in Vienna: returning to a plasma-physics working meeting after twelve years, and finding that AI may help with the reasoning as well as the computation