Fork-and-Flush: escaping autoresearch idea basins by resetting chat context
RobertTLange · x · 2026-10-27
Robert Lange summarizes a new approach to AI autoresearch. Experiments show 8 independent agents converge into 5 distinct "idea basins" on the same task, stuck even with more compute. Fork-and-Flush forks the agent into K branches that keep the workspace but reset chat history, runs each on part of the budget, keeps the best branch, and repeats — basin hopping plus population-based training applied to agents, where dropping chat history acts as the perturbation that lets a fresh-context probe jump into a new basin.
More from coding & agent
- Sentient OS launches: on-device AI that moves your cursor and drafts replies in your voice — thisiskp_ · 2026-10-28
- Grok Bot downloads straight from Cursor's servers, suggesting xAI built it on Cursor's stack — faraz0x · 2026-10-28
- 90% of a SWE career is complaining about bad codebases — until coding agents arrived — vivekhaldar · 2026-10-28
- Radiology journal publishes review on agentic AI systems, topologies and evaluation — jnkath · 2026-10-28
- Parsers launches: turn URLs from Google Maps, TikTok, G2 into clean JSON via one API — gaganghotra_ · 2026-10-28
- The AI CEO experiment failed: same-model agent teams can't truly diversify — thisguyknowsai · 2026-10-28