Task horizon

A measure of how much autonomous work an AI agent can perform over a continuous span of time — in effect, the length and complexity of tasks an agent can execute end-to-end without requiring human intervention.

Description

Lance Martin of Anthropic uses task horizon as a metric for tracking the capability progression of successive Claude models, plotting each model against the amount of autonomous work it can sustain over time. → The concept frames model improvement not merely in terms of benchmark scores or single-turn quality, but in terms of how far an agent can operate independently before needing guidance, correction, or a human handoff.

Significance for agent engineering

Task horizon is a practical lens for understanding why certain agent architectures, evals, and reliability requirements become necessary as models grow more capable. A short task horizon implies the agent will frequently return control to a human; a longer horizon means errors can compound over many steps before they are caught, raising the stakes for robustness, tool use, and planning. Tools such as Claude Code are implicitly designed to push this horizon outward by giving models reliable access to execution environments across extended coding and reasoning sessions.