On Oct. 19, 2026, jasoncrawford asked in a post on Twitter whether anyone had put roughly 10,000 agents to work on fusion reactor design, particularly a stellarator. The post suggested the task might suit agents because it had clear metrics and, in the author's view, could be iterated on entirely in simulation. It posed a question rather than describing an experiment.
In a quote-post of jasoncrawford, Andercot offered to set up an open-source agentic sandbox if a frontier lab sponsored it. Andercot suggested calling it “Fusion Bench” and comparing AI models, agent teams or agentic harnesses. The post pitched 1 cent per MWh electricity as a benchmark quest, not a measured result.
jasoncrawford’s question also drew cautions about what agents could contribute. DanielleFong replied that agents had been used “somewhat, but only on certain corners” of what she called “our reactor.” She warned that agents need close watching because they can bring assumptions into domains where they do not work. sophsmeadowx2 questioned what 10,000 agents would add over gradient descent on the same objective and raised superconducting coils and materials as challenges.
Quote-posting Andercot’s proposal, DavinYCD argued that each design should disclose the assumptions behind its electricity-cost estimate. Different assumed plant lifetimes or operating hours, DavinYCD warned, could cause a leaderboard to rank accounting choices alongside reactor designs. In another reply to jasoncrawford, ramez said there had been substantial machine learning in fusion reactor design but that he was not aware of any use of large language models, though he imagined there might have been some. None of these replies identified a roughly 10,000-agent reactor-design experiment.
