Long-horizon tasks — those requiring hundreds of steps over hours or days — are the current frontier of agent capability. Error compounds; context grows unmanageable; state must be externalized. Almost every honest benchmark shows sharp capability drops past ~20 steps.
- 01Explain why error compounds in long horizons.
- 02State the three mitigations that reliably help.
- 03Name a long-horizon benchmark.
The compounding problem
If each step succeeds with 95% probability, a 20-step task completes 36% of the time; a 100-step task, 0.6%. Long-horizon reliability requires per-step reliability much closer to 100% than most current agents achieve.
Mitigations
1. Externalize state (write to disk, track a todo list). 2. Checkpoint and verify (each phase validated before proceeding). 3. Human handoff points (agent stops and asks). These do not solve the problem; they make failures visible and recoverable.
Benchmarks
SWE-bench (software engineering), OSWorld (computer use), WebArena (web navigation). All show sharp capability cliffs on tasks longer than a few steps.
- Error compounds multiplicatively in long horizons.
- Externalized state + checkpoints + human handoffs are the reliable pattern.