Plan Validate Execute

Decouple plan generation from execution so a bad plan can be rejected before it burns tool calls, tokens, and irreversible side effects.

A task has a goal and constraints; a plan is the roadmap of steps to satisfy them. Coupling planning and execution in one prompt (think-then-act in a single pass) can let an unvalidated runaway plan consume time and money before anyone notices.

Prefer three stages:

  1. Plan generation (task decomposition into actions).
  2. Plan validation — heuristics (tools must exist in the tool inventory; length caps) and/or an AI judge that regenerates unreasonable plans.
  3. Execution — usually via function calling.

Plans need not cover the whole task; they can address a subtask. Hierarchical planning further separates high-level roadmap from detailed expansion. Splitting generation, validation, and execution into separate roles is already a simple multi-agent architecture. Parallel candidate plans traded against an evaluator buy latency at higher cost.

Sources

  • AI Engineering (Chip Huyen), ch. 6