Harness²: Google's self-improving agent framework adds +34.7 pts without training
Prompt Engineering · youtube · 2026-10-24
Harness², from Google Cloud AI Research and Northwestern, lets an AI agent rewrite its own harness (prompts, skills, scripts, subagents, control-loop hooks) for each task without retraining. The video covers:
- Why one-size-fits-all harnesses are a bottleneck and how agents can evolve them without an answer key
- The Propose–Probe–Compose validation loop with two levels of recursion
- An experience library that accumulates reusable harness improvements across tasks
- Results: Harness² Evolved scores 77.8 vs 43.1 for a default harness and 60.4 for single-layer self-editing (+34.7 pts), with comparisons against Claude Code and Codex, plus where the approach breaks
Paper: arxiv.org/abs/2610.14224; open-sourced at google-research/harness2.
More from coding & agent
- Microsoft open-sources Agensh, a decentralized multi-agent harness scaling to 1,024 agents — donglixp · 2026-10-24
- dumbtv strips Fire TV ads and tracking with one command, no root needed — Signalman23 · 2026-10-24
- CodeRabbit CLI v0.9.0 ships cr code for managing cloud agent tasks in terminal — leslysandra · 2026-10-24
- Solo dev turns a school timetable solver into an MCP server for Claude — Wilbek_DK · 2026-10-24
- ChatGPT Sites ships detailed site analytics: dev logs +216% weekly traffic — Dimillian · 2026-10-24
- Gemini Flash-Lite as a meeting copilot brain: real test results, what worked and what broke — DeskSignalAI · 2026-10-24