Autoresearch agents get stuck in local optima — the fix may be making them forget
eshorten300 · x · 2026-11-03
On the 149th Weaviate Podcast, UC Berkeley PhD student Shangyin Tan explains why long-running autoresearch agents stall like numeric optimizers: as the history of attempts piles up in context, the agent anchors on it and every new idea becomes a small tweak of what already exists.
The fix is simple: throw the history away. Keep the best solution, drop the path that led to it, and let the agent explore without conditioning on extra context.
The work builds on GEPA and its successor optimizeanything, extending these ideas from prompt optimization to code, GPU kernels, and agent harnesses. Tan also draws a sharp distinction: prompts and harnesses are generalization tests, while a single kernel is pure optimization.
More from coding & agent
- Dev builds visual layout editor for MOSS-Image that outputs JSON prompts — jambavant · 2026-11-03
- Sentry CEO shares coding agent asking approval to rewrite Sentry in Rust — zeeg · 2026-11-03
- Microsoft shaved 4.5% off clarity.js with 50+ AI agent experiments — gaganghotra_ · 2026-11-03
- Vibe-devops agents may rebuild SaaS in-house as vendors like Bending Spoons triple prices — robleclerc · 2026-11-03
- rauchg: stateless sandboxes are the wrong benchmark as Vercel Sandbox gets 2.4x faster — cramforce · 2026-11-03
- Model swap cut costs 59% and latency 75% — and he still blocked the release — Fun_Employment6042 · 2026-11-03