Several months ago we published a Verilog simulator built by AI agents in 43 days. The project is still running, with more than 10 agents writing simultaneously to a single shared branch. There are no PRs and no separate worktrees. The agents constantly bump into each other, and velocity holds anyway, because the agents resolving the conflicts are more capable than the conflicts themselves.
We treat coordination as a temperature parameter. The project runs hot early, while the architecture is in flux and inconsistency is cheap, then anneals as the system matures and bugs take hours to trace. We call this thermodynamic programming, and the same structure appears in neural network training and collaborative writing. Full piece here: normalcomputing.com/blog/thermodyn…
English
