Growing Harness moves recurring control into code learned from failure traces, cutting LLM calls 76-92%
arXiv 2609.26760·medium signal
The method starts from a scaffold with no strategy built in. It localizes each failure to a function-level code surface, repairs a window of failures together, and rolls back any repair that fails a held-out gate. Accepted edits accumulate in one shared harness. On WebArena-Verified the grown harness held 44.7-45.3% success from 4B to 120B models, while a plain tool-calling agent fell to 6.7% at 4B. Deployed inference cost fell 74-99%. The practical pattern: when an agent keeps re-deriving the same control flow, promote it to code behind a regression gate.