The Compound Mistake Problem

Question 1

What is the Compound Mistake Problem?

Answer 1

In a multi-step agent workflow, errors accumulate exponentially: Accuracy_total = (Accuracy_step)^n. A 95%-per-step agent on a 10-step task has only ~60% overall success. A 50-step task drops to nearly zero.

Question 2

What agent design principles follow from the Compound Mistake Problem?

Answer 2

(1) Avoid monolithic “do everything” agents; (2) Design systems as chains of specialized, smaller agents; (3) Use error-correction loops where agents reflect on outputs before proceeding; (4) Keep chains as short as possible.

Question 3

What is the Planner-Evaluator-Executor pattern?

Answer 3

A decoupled agent architecture: (1) Planner LLM generates proposed actions; (2) Evaluator (deterministic policy or judge model) reviews the plan for safety, compliance, and correctness; (3) Executor carries out only approved actions. Supports human-in-the-loop and Policy-as-Code.

Question 4

Why is the Planner-Evaluator-Executor pattern important for federal AI?

Answer 4

It enables Policy-as-Code enforcement — deterministic rules can block unsafe actions regardless of what the probabilistic model generates. For High-Impact AI (OMB M-25-21), the Evaluator step can include human review of the plan before execution.

Question 5

What is a “mixture of agents” approach?

Answer 5

Designing agentic systems as chains of specialized, smaller agents rather than one large monolithic agent. Each agent handles a focused sub-task, reducing the number of steps per agent chain and mitigating the Compound Mistake Problem.