Google paper: agent harnesses evolved on one model underperform when moved to another
blaizedsouza · x · 2026-11-19
A Google paper on Continual Recursive Self-Improvement (CRSI) shows an agent harness evolved on Claude Opus scores below Gemini's plain baseline when transferred. Their fix, Self-Improvement Memory (SIM), logs each edit's rationale and outcomes, lifting final scores by an average of 2.36 points across ten coding source runs (p=0.003).
More from coding & agent
- 10 Free Repos That Teach AI Agent Harness Engineering: Tools, Memory, Rules — blaizedsouza · 2026-11-19
- Kilo adds linked billing with Claude Platform, no API key sharing needed — anacondainc · 2026-11-19
- Stanford team introduces DiscoveryAgent for rigorous open-ended scientific discovery — james_y_zou · 2026-11-19
- Solo dev cuts sandbox cold-start to 0.45s, sleeping VMs cost 1¢/month — emilahlback · 2026-11-19
- Team cuts agent sandbox cost 88% by moving to GKE, open-sources gke-sandboxes — emilahlback · 2026-11-19
- Google Antigravity shipped 150 releases in 90 days — a Googler's hands-on review finds it finally caught up — rseroter · 2026-11-19