Exponential Progress: AI Agents Doubling Task Complexity Every 7 Months, METR Research Finds
Key Takeaways
- ▸AI agent task completion capacity is doubling every 7 months (recently accelerating to 4-month doublings), representing consistent exponential progress across all frontier systems
- ▸Task complexity has grown from 30-second operations at ChatGPT's 2022 launch to 14+ hour autonomous tasks by 2026
- ▸Current extrapolations predict month-long task automation by 2027-2028 if trends hold, with potential for even faster acceleration
Summary
Researchers at METR have discovered a new exponential trend in AI agent capabilities—analogous to Moore's Law for semiconductors—revealing that the maximum task duration frontier AI systems can autonomously handle is doubling approximately every 7 months. The finding is based on testing agents from 2019 to 2026 across roughly 230 tasks, primarily coding challenges but also general reasoning problems, establishing a strong correlation (R² = 0.83) between task complexity and agent success rates.
The "time horizon" at which agents achieve 50% success is growing exponentially. In 2022, when ChatGPT launched, agents could handle 30-second tasks; today's frontier systems autonomously complete coding tasks requiring over 14 hours of human effort. Extrapolating this trend, METR projects agents could handle full work-day tasks (8 hours) by 2027, week-long tasks (40 hours) by 2028, and month-long tasks (167+ hours) by 2029.
Notably, the trend has recently accelerated—doubling every 4 months in 2024-2025 rather than the historical 7-month cadence. Researchers speculate that as AI systems improve, they could trigger a flywheel effect where capable agents accelerate development of even more capable agents, potentially creating superexponential growth in capabilities. The team frames this trajectory as potentially "one of the most important trends in human history."
- Researchers warn of a potential recursive flywheel where improved agents directly enable building more capable agents, risking superexponential growth
- The trend has major implications for workforce automation, AI research acceleration, and the timeline to transformative AI capabilities
Editorial Opinion
This research represents one of the most consequential findings in AI development—not because it reveals a breakthrough in any single model, but because it documents a consistent, accelerating trend in agent autonomy that cuts across all frontier systems. The exponential doubling every 7 months (now 4 months) is remarkable; what's more striking is the speculative scenario of recursive acceleration, where improved agents directly enable building even better agents. If this flywheel activates, we could see capabilities that surpass human expertise across research, engineering, and strategic domains within 18-24 months. The researchers are right to flag this as potentially transformative—the workforce implications alone warrant immediate attention from policymakers and industry leaders.
