Definition
The time horizon of an AI agent is the length of tasks, measured in how long a human needs for them, that the agent completes with 50% success. METR tracks it across models.
The time horizon doubles roughly every 7 months: seconds in 2019, hours in 2026. If the trend continues, week-long tasks are 2 to 4 years away. Longer horizons mean more autonomous steps without a human in the loop.
Appears in
- Lecture 1, METR: AI Task Time Horizons
- Lecture 8, The Agent Capability Stack: what agents can do
- Lecture 13, METR: ~4-month doubling since 2023, frontier 12 to 17 h
- Lecture 13, domains: GUI use and driving lag