Agents stop after a median 52 minutes, but longer runs score higher 62% of the time
teortaxesTex · x · 2026-11-14
Dominik Scherm's agent experiments: given a 6-hour budget, the median agent run stops after just 52 minutes. Yet persistence pays — when a model spent at least 2x longer on a task than in another attempt, it scored higher 62% of the time, by 7 points on average. Tasks mixed public and proprietary lab data. Relevant to agent eval design and how to incentivize longer-horizon work.
More from coding & agent
- Hands-on DeepSeek's $0.06/hr ASR API: loses to a locally fine-tuned 1.7B model — AGI Hunt · 2026-11-14
- SAG Wiki: open-source local knowledge base that skips embeddings and vector DBs, works with Claude Code — FellMentKE · 2026-11-14
- Recursive self-improving agent MolRSI lifts wet-lab hit rate 82%, beats ClaudeScience 8-0 — 机器之心 · 2026-11-14
- OpenAI $200 plan drops to $100 on Nov 20; user races to burn 4 banked Codex resets — koltregaskes · 2026-11-14
- Tonbi ships AI Garage plugin giving agents MCP access to his content catalog — Teknium · 2026-11-14
- Google's Modern Web Guidance lifts coding agents' modern web compliance from 56% to 95% — aigclink · 2026-11-14