OpenAI data: zero-intervention agent success hits 79% on run monitoring but drops for algorithm implementation
daveholtz · x · 2026-11-19
OpenAI's report breaks down autonomous agent success rates by estimated human task duration: success on long-running tasks keeps improving over time (e.g., 4–8h tasks from 12% in March to 35% by October), though this reflects both capability gains and shifts in task mix.
Using Epoch's ONET for AI R&D taxonomy, success varies widely: zero-intervention success is 79% for monitoring runs for issues and 75% for answering technical questions, but falls below 30% for implementing research algorithms and building training infrastructure.
More from AGI Musings
- OpenAI data shows human time on safety and alignment up 77.2% in 2026, fastest-growing research area — daveholtz · 2026-11-19
- OpenAI: as AI improves itself, governance challenges grow — humanity must stay in the loop — daveholtz · 2026-11-19
- OpenAI: researcher code output now 7x pre-2025 average, OECI frontier rising 16.4 points/year — daveholtz · 2026-11-19
- OpenAI researchers would trade 40% of team headcount to keep frontier coding agents — daveholtz · 2026-11-19
- OpenAI: median researcher spends $2,400/day on agents, token costs now 2x labor costs — daveholtz · 2026-11-19
- OpenAI report calls for labs to publicly track recursive self-improvement as AI use in research soars — OpenAI · 2026-11-19