Overview
This page is the atomic definition. Multi-agent coordination patterns live at multi-agent.
Definition
The planner-executor pattern is an agent architecture that separates reasoning into two distinct roles. The planner receives the high-level task and produces a structured plan, typically a numbered list of steps, without executing any of them. The executor receives one step at a time and carries it out, often via tool-call invocations, without needing to reason about the overall strategy. The separation reduces cognitive load on each model call: the planner is not distracted by execution details; the executor is not distracted by strategic choices. The two roles may be the same model invoked sequentially with different prompts, or different models (a more capable model planning, a cheaper model executing). Re-planning is triggered when the executor encounters an unexpected result that invalidates subsequent steps. Dynamic replanning keeps the architecture robust to real-world surprises but adds latency and cost. This pattern is used in ReAct (Reason + Act) and Plan-and-Execute frameworks.
When it applies
Use planner-executor for: complex multi-step tasks with more than 5 steps, tasks where early failures should trigger plan revision, and tasks where parallel execution of independent steps is desirable. For simple 2-3 step tasks, a flat agent-loop without explicit planning is sufficient.
Example
Task: “Research competitors and draft a summary.” Planner outputs: 1. Search for competitor A. 2. Search for competitor B. 3. Read A’s pricing page. 4. Read B’s pricing page. 5. Write comparison. Executor runs each step sequentially with tool calls. If step 3 fails (page not found), the planner is called again with the partial results to revise steps 4 and 5.
Related concepts
- agent-loop - the execution cycle the executor runs within; planner-executor is a structured form of the agent loop.
- tool-call - the executor emits tool calls to carry out each plan step.
- structured-prompt - the planner typically outputs a structured format (JSON or numbered list) consumed by the executor.
- system-message - separate system messages define planner versus executor roles.
- multi-agent - planner-executor is a two-agent topology within a multi-agent system.
Citing this term
See Planner-Executor (llmbestpractices.com/glossary/planner-executor).