Hashimoto's rule for AI agents is that every mistake becomes a permanent fix

An X post summarising the Harness Engineering playbook. Names the practice "Hashimoto's rule": every mistake becomes a permanent fix, living in a rule file, a test or a permission boundary rather than a prompt tweak. Repeats the playbook's harness-only results (LangChain's Terminal Bench climb, an OpenAI team's 1,500 pull requests, a 31% to 75% GAIA move with the model fixed) and its tell for a healthy system: rule-file growth fast at first, then about one rule a week. Draws the line at one-off work, where the harness is not worth its cost.