One Bad Step, N Bad Steps: How LLM Agent Failures Cascade

Many developers who reason about AI agents statistically often misjudge how errors behave. They measure a per-step error rate—such as 10% of steps producing incorrect outputs—and assume errors are entirely independent. Under this assumption, a single wrong step stays contained while the rest of the run proceeds normally. However, observing a real execution trajectory reveals a starkly different reality.

In practice, if step 4 gets a fact slightly off, step 5 reasons on top of that erroneous premise and commits to it even harder. Step 6 then takes an action based on both accumulated flaws, and by step 8, the agent is executing a doomed plan with absolute confidence. One mistake transforms into five because the errors are coupled through the context, turning a modest 10% step-error rate into a severely corrupted run.

This phenomenon is known as a cascade: a single fault amplifying vertically down a trajectory. Unlike the horizontal spread seen in distributed retry patterns where a bad condition hits multiple workers simultaneously, a cascade spreads through time within a single run because an agent's past output serves as its future input.

<

A stateless function that fails simply returns an error. An agent that fails does something far more dangerous: it writes the failure down into its transcript where it can read it again. The exact mechanism that makes agents functional—accumulating transcripts where every step conditions on prior history—is the exact channel down which mistakes travel.