Step fourteen: why long AI agent runs fall apart

About this episode

How errors build up in long-running AI agents: what context poisoning is, the maths of compounding errors over many steps, and the common failure modes, from hallucinations and step repetition to premature termination. Then what helps: clear tool errors, verification steps, small tasks, handoffs, retry and step limits, and human review.

Written by Kiran Krishna. Narrated with an AI voice. · LinkedIn

In this episode

1. 0:01In 2025, Anthropic let Claude run a small shop in its own office, for about a month.
2. 1:03Here's the plan.
3. 1:14So what actually happens at step fourteen?
4. 3:53So how likely is a bad step like that, somewhere in a long run?
5. 5:48So does this happen in real runs?
6. 8:02So what actually helps?
7. 10:12So why do long runs fall apart, when every step looks fine?