Google paper: evolving the agent harness doubles GPT-5-mini's SWE-bench score to 48.3%
bittingthembits · x · 2026-10-13
A new Google paper (Evo²) shows the harness around an agent — tools, memory, workflow — matters as much as the model. One AI plays two roles: a Proposer edits its own harness, a Reflector diagnoses each edit; edits and results become training data via "Studying the Textbook" and "Scrimmage Practice." Same GPT-5-mini, better harness: 24.4% → 48.3% on a SWE-bench Verified subset (strongest baseline 41.0%), 82.1% vs 62.6% on Polyglot, with evolved harnesses transferring across benchmarks. Note: the tweet also shills a $TAO token; verify against the paper.
Related event: DeepMind Paper Evolves Agent Harness Together with Weights(2 posts)→
More from coding & agent
- Dev builds MCP layer so Claude Code can work on servers without ever holding SSH keys — NETMESAFE · 2026-10-16
- As coding agents run longer autonomously, one dev's tip: don't parallel-session, just breathe — teropa · 2026-10-16
- 'Refusing to code with AI in 2026 is like refusing Google in 2005' — prasenx · 2026-10-16
- Claude Code adds idle compaction to save 1h prompt cache on idle sessions — EricBuess · 2026-10-16
- Handling agents that crash mid-task: the hard part is the unknown outcome across the network — Responsible-Dare5139 · 2026-10-16
- OpenAI DevDay Community Hackathon heads to Berlin with developer, founder and GTM tracks — paw_lean · 2026-10-16