Why do long-running agents fail?

Long-running agents fail because a small chance of error at each step compounds over a run that never ends, and the agent keeps acting on its own record of what already happened, so one bad assumption carries forward into every later action. Nothing in that kind of run forces a reset, so the mistake surfaces late.

In METR’s Time Horizon 1.1 results, last updated September 8, 2026, Anthropic’s Claude Opus 4.6 is predicted to succeed half the time on software tasks that take a human expert about 12 hours (METR’s confidence interval runs from about 5 to 61 hours), but 80% of the time only on tasks that take about 70 minutes. An always-on agent strings together far more work than either, so its week depends on how often each piece goes right, the multiplication behind why success rates drop on longer tasks.

Context is the other pressure. A history that never ends outgrows any context window, and at the limit the call either fails outright or silently loses its oldest turns. A bigger window doesn’t fix it on its own: in Andon Labs’ February 2025 Vending-Bench paper, agents running OpenAI’s GPT-4o mini with 10,000, 30,000, or 60,000 tokens of memory did worse with more memory, not better.

METR says its measurements above 16 hours are unreliable with its current task suite, and an always-on agent has no end point at all, so METR’s numbers describe the pieces of its work, not the whole run.