Google paper shows distilling the agent harness lifts task success from 23.3% to 44.3%
omarsar0 · x · 2026-09-24
A new paper from Google and colleagues studies whether an agent harness can be distilled into the model itself.
- With the specialized harness removed, macro task success jumps from 23.3% to 44.3% — above the base model's 41.7% even with the harness attached.
- Harness-Zero uses the optimized harness only during training: since its action space differs from the deployment harness, a harnessing agent corrects the student's responses before they run, and those corrected runs become training demonstrations.
- Across 28 harness-induced behaviors in knowledge work, tool use, and science, 82.3% are recovered on average.
- For frontier models using the same evolved harness, the agent-as-harness form also beats code-as-harness.
Robustness remains to be seen, but the approach is notably interesting.
More from coding & agent
- GitHub ships new maintainer controls for agentic coding, keynote slides open-sourced — marlene_zw · 2026-09-24
- Next.js pre-announces Sept 30 security release fixing 9 vulnerabilities — cramforce · 2026-09-24
- The agent era needs goal description, not task instructions — manosaie · 2026-09-24
- Open-source WorkSwarm keeps 5-person shared agent coherent over 189 turns, lifts SWE-bench Lite to 87% — HeyZoyaKhan · 2026-09-24
- Developer built 44 proprietary tools: the right harness pushes agents far beyond raw models — D3VAUX · 2026-09-24
- Open source project reverse-engineers every retro game with AI agents, starting with C64 — Vjeux · 2026-09-24