An autonomous AI agent task horizon can be measured in elapsed time, number of dependent steps, task complexity, or the amount of state an agent must preserve. A longer horizon means the system can pursue goals through more observations and actions before supervision is required.
Horizon claims must include reliability and task difficulty. Agents may run for hours while making little progress, and small error rates compound over many steps, so evaluation should verify accepted outcomes, unsafe actions, recovery, and the amount of human correction.
ELI5
An autonomous AI agent task horizon describes how far an agent can get through a job on its own before a person must correct or redirect it. The horizon can be measured by time, number of connected steps, difficulty, or how much changing information the agent must keep straight.
For example, an agent may reliably complete a five-step data cleanup but begin making mistakes during a fifty-step migration. Simply running for many hours does not prove a long horizon, because useful progress, correct outcomes, recovery, and unsafe actions all matter.
