Google Open-Sources Harness²: Agents Rewrite Their Own Scaffolding, Weights Frozen, +11.5 Points
Arindam_1729 · x · 2026-10-19
Google open-sourced Harness², a framework that improves AI agents without swapping models or fine-tuning: with weights frozen, the agent rewrites its own harness — prompts, tools, memory, and control flow.
Reported gains: WorkBuddy-Bench +11.5 points, JobBench +4.5, LAB +11.8 percentage points — same model, better scaffolding, real jumps.
How it works:
- Task-level recursion: the agent proposes harness changes, runs them, compares behavioral differences across versions, and merges the best ones — no ground-truth answers needed.
- Harness-level recursion: an experience reflector stores harnesses that worked plus evidence-backed lessons, carrying improvements forward across tasks.
The framework can modify 8 components: system prompts, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins — essentially the full editable surface of a modern agent.
Key takeaway: when agent results stall, the bottleneck is often the harness, not the model. The next leap may come from agents that learn to build better systems around themselves.
More from coding & agent
- Qwen Code Desktop v0.25.3 Ships With Batch of Managed-Agent Fixes — github-actions[bot] · 2026-10-19
- Teknium to Turn Hermes Agent Into a DIY Game Dev Studio — Teknium · 2026-10-19
- Agent Control Plane explained: John Savill's quick overview clears up common confusion — JohnSavill · 2026-10-19
- Microsoft's ASSERT and ACS: blocking AI agents from accessing the wrong account — usamawahabkhan · 2026-10-19
- Exa Snapshot lets agents search the web as it looked on any past date — AI Engineer · 2026-10-19
- Anthropic Researcher: Claude Wasn't Worth Using at 80% Accuracy, Now Writes Most Code at 99% — victor_explore · 2026-10-19