Skip to main content

Plan-and-Execute

Separates reasoning into two distinct phases: a planner creates a high-level plan upfront, and an executor works through it step by step. If a step fails or new information emerges, the plan can be revised. This prevents the agent from losing sight of the overall goal while deep in execution.

Structure​

The planner sees the full goal. The executor focuses on one step at a time. If execution reveals the plan is wrong, control returns to the planner for revision.

How It Works​

  1. Plan — planner LLM analyzes the goal and produces an ordered list of steps
  2. Execute — executor works through steps one at a time (often using ReAct per step)
  3. Track — progress is tracked against the plan (completed, in-progress, pending)
  4. Replan — if a step fails or produces unexpected results, the planner revises the remaining steps
  5. Complete — when all steps are done, results are assembled into the final output

The planner and executor can be:

  • The same model with different prompts
  • Different models (cheap model for planning, capable model for execution)
  • Different agents with different tool access

Key Characteristics​

  • Goal-oriented — the plan keeps the agent focused on the overall objective
  • Recoverable — replanning handles unexpected situations without starting over
  • Separation of concerns — planning and execution are independent cognitive tasks
  • Overhead — planning step adds latency before any execution begins
  • Plan quality matters — bad plans lead to wasted execution cycles

When to Use​

  • Long-horizon tasks with many steps (research, multi-file code changes, data pipelines)
  • The agent tends to lose track of the overall goal during execution
  • Tasks benefit from upfront decomposition before diving into details
  • You need progress tracking against a plan
  • Failures in one step should trigger intelligent replanning, not just retries