Auto‑train the harness, not the LLM, for cross‑model gains
The Harness‑Training repo implements a PyTorch‑style trainer that freezes the LLM and repeatedly optimizes the surrounding harness—prompts, context handling, tool calls, and repair loop. Each epoch proposes a git‑style diff to the harness, runs a panel of benchmarks, and automatically promotes or rejects the change, delivering cross‑model performance gains.