Overview

The Planner-Generator-Evaluator pattern is an agentic design pattern for LLM-assisted software development. It decomposes a development workflow into three specialised agent roles with distinct scopes and responsibilities, preventing any single agent from holding too much context or making unreviewable decisions in one pass.

The three roles

Planner

The Planner’s scope restriction is deliberate: keeping it at spec-level means its output is reviewable by a human before any code is generated.

Generator

The incremental approach limits blast radius: if a sprint’s output is wrong, the damage is bounded and the Evaluator can catch it before the next sprint begins.

Evaluator

The Evaluator grading against hard thresholds is the quality gate: it prevents Generator drift from accumulating across sprints undetected.

Why this pattern matters

Single-agent coding loops have two failure modes: (1) the agent drifts from the original intent across a long context window, and (2) the agent self-evaluates optimistically. The Planner-Generator-Evaluator pattern addresses both:

The pattern mirrors human engineering team structure: architect (Planner), engineers (Generator), QA (Evaluator).

Resources