Topic 6

Compound errors

AI Engineering (Ch. 6, "Agent Overview") gives two reasons agents need stronger models than single calls. One is higher stakes: with tools, failures can have more severe consequences. The other is compound mistakes: an agent takes many steps, and accuracy falls as steps add up.

Predict first

Each step of an agent is right 95% of the time. Roughly how often does a 10-step task go right from start to finish?

95%About 85%About 60%About 10%

Per-step accuracy, compounded

Arithmetic · assumes every step must succeed and failures are independent

Accuracy per step: 95.0%Steps: 10

Whole task succeeds

59.9%

At least one step fails

40.1%

1 steps95.0%
5 steps77.4%
10 steps59.9%
20 steps35.8%
50 steps7.7%
100 steps0.6%

Real steps aren't independent, and an agent that notices a mistake can recover from it. The point stands: long tasks need very reliable steps, or checks between them.

The calculation assumes every step must succeed and failures are independent. Real agents are messier both ways. Steps aren't independent, and an agent that notices a mistake from a tool result can recover, which is what reflection is for. But the shape is why long tasks are hard: at 99% per step, 50 steps still come out at about 61%.

That suggests three levers, and this module covers each:

Fewer model decisions. Steps that don't need judgment can be fixed in code (lesson 2).
Checks between steps. Validate arguments, check results, and catch errors before they compound (lessons 4 and 5).
Shorter leashes. Limit how many steps an agent can take, and stop to ask a person before the risky ones (lesson 6).

Quick check

Your agent is 98% reliable per step and tasks take about 30 steps. What's the most effective change?

Accept about 55% task successMove steps that don't need judgment into fixed code and add checks after risky steps, so fewer steps can fail silentlyAdd more toolsRaise the temperature