A new Moore’s Law for AI agents

Abstract

An interactive explainer presenting METR’s finding that the length of coding tasks AI agents can complete autonomously is growing exponentially. Testing frontier systems from 2019 to 2026 on roughly 230 tasks, mostly coding with some general reasoning, METR found that task length is highly correlated with agent success rate (R² = 0.83) and that the “time horizon” — the task length an agent completes 50% of the time — has doubled roughly every seven months. Over 2024–2025 the doubling time fell to about four months. Extrapolating the slower trend implies agents handling one-work-day tasks in 2027, one-work-week tasks in 2028 and one-work-month tasks in 2029; under the accelerated trend, month-long tasks could arrive as early as 2027. The piece argues that growth may become superexponential as AI systems increasingly help build their successors, and that if this feedback loop automates AI research, the rising time horizon could prove one of the most consequential trends in human history. – AI-generated abstract.