Harness-Zero Distills Agent Harness Gains into Model Weights, 23.3% to 44.3%
Haoran Ye · hf · 2026-09-22
Agent harnesses — the external systems mediating model-environment interaction — boost agent performance, but the best harness varies by domain, instance, and model. Harness-Zero performs harness distillation: a domain-optimized harness guides training, and the behaviors it induces are transferred into model weights so gains persist under a single fixed deployment harness. The key trick is agent-as-harness: a harnessing agent corrects student responses in the target harness's action space before execution, turning harness guidance into training demonstrations.
Findings:
- For frontier LLMs with the same evolved harness, agent-as-harness outperforms code-as-harness
- With the specialized harness removed at deployment, the base model's macro-average success rises from 23.3% to 44.3% — even beating the 41.7% achieved with the harness still attached
- Across 28 behavioral patterns in knowledge work, tool use, and science domains, it recovers 82.3% of harness-induced behaviors absent from the base model
More from coding & agent
- Halo open-source post-training framework hits 2.8x TRL throughput — prasenx · 2026-09-22
- Jev from a ChatGPT co-creator auto-routes tasks to the cheapest capable model — fsharpman · 2026-09-22
- Prof. Tom Yeh releases printable hand-calc agentic AI cost math problems — ProfTomYeh · 2026-09-22
- apibase unifies 327 tools from 92 providers behind one pay-per-call MCP endpoint — modelcontextprotocol · 2026-09-22
- mcp-server-s3 ships MCP server letting agents browse, upload and share S3 files — modelcontextprotocol · 2026-09-22
- Dev built an AI game-playing plugin but shelved it: vision models too slow and costly — ezshine · 2026-09-22