Google & JHU's Pro² Agent hits 76% MLE-Bench medal rate by tracking bottlenecks
rohanpaul_ai · x · 2026-11-21
A Google–Johns Hopkins paper introduces Pro² Agent, which shifts an AI research agent's primary state from the codebase to an explicit 'Problem State': suspected bottlenecks, candidate fixes from the literature, and confidence in each, updated via single-variable experiments before committing changes. It reaches a reported-best 76.0% medal rate on all 75 MLE-Bench tasks within 48 hours (70.7% at the standard 24-hour budget), and its first solution already medals on 57.6% of MLE-Bench Lite vs 28.8% for MARS. Takeaway for agent builders: write down and update what is blocking progress, and test one change per experiment.
More from coding & agent
- Reverse-engineering Apple Watch: a single Claude Code prompt builds a physics watch face in 30 minutes — mtizard · 2026-11-21
- us-legal-mcp: an MCP server exposing comprehensive US legislation to AI agents — modelcontextprotocol · 2026-11-21
- Maindex launches Expert Memory Service, a full-control memory API for AI agents — modelcontextprotocol · 2026-11-21
- AI-built oil painting engine: 2,300 commits in 8 days as Claude Code runs a 'studio' of sessions — FotoRe_store · 2026-11-21
- Open-source chirp gives your AI agents a shared, self-editable message board — itsarnavb · 2026-11-21
- Frontier agents complete under 8% of tasks in 26-person simulated org, study finds — vasuman · 2026-11-21