Harness Co-Training: Models shaped inside agent loops become new industry norm

DynamicWebPaige · x · 2026-08-18

The article analyzes the trend of "Harness Co-Training," where models are trained within agent scaffolds before release. It contrasts strategies from poolside, Cursor, Z.ai, and Anthropic: poolside trains across multiple harnesses on purpose, Cursor injects corrections at tool failure points, Z.ai reports best-harness scores, while Anthropic argues for leaving the model untouched.

Commercially, model capabilities become tied to specific scaffolds, meaning leaderboard scores reflect a "model-plus-harness" pair. Switching environments may degrade performance.

Original post →

More from coding & agent

coding & agent channel →