The harness writes itself

The scaffolding around a model — its tools, prompts and control flow — had been the last reliably human-authored layer. When the agent mines its own failures and rewrites that layer, improvement closes a loop that needs no engineer. The thread starts small, with users metaprompting Codex to write its own objective file in May, and is named for 25 June 2026, when Self-Harness had an agent mine its weaknesses and rewrite its own scaffolding for double-digit Terminal-Bench gains. In July a packaged skill for meta-harnesses shipped with the warning that the loop optimizes whatever signal you give it, Weco’s outer loop rewrote its inner researcher seven times, and Kimi K3 spent 17 hours rewriting its own harness for eleven points. In August Qwen3.8-Max, left alone sixteen days, shipped 265 commits and its own self-evolving harness, and Poetiq’s self-optimizing optimizer claimed state of the art by upgrading harnesses rather than weights; by September a model installed its own tooling on the way to a task. It is the harness sub-thread of recursive self-improvement.