29. New research: long-running agents often fail by stopping too early, not because the model can't make progress. by @ii_posts (Intelligent Internet) · backlist 2026-05-08 · rubric 96.0