Planner

A planner is the part of an agent system responsible for turning a goal into a plan. In simple agents this is the LLM itself, prompted to reason step by step. In more sophisticated systems the planner is a dedicated component, sometimes its own model, that produces a structured plan (a list of steps with dependencies) that another component (the executor) carries out. The plan-and-execute pattern, popularized by LangGraph and Anthropic's research, separates these roles to keep the executor focused on action while the planner focuses on strategy. This separation also makes the plan auditable before execution begins.

Two design tradeoffs matter. First, plan-then-execute (full plan upfront) gives you auditability and human-approval checkpoints, but the plan goes stale if early steps change the world. Second, interleaved plan-and-execute (revise the plan after each step) is robust to changing world state, but harder to audit and more expensive in tokens. Production systems often start with full upfront plans for short workflows and switch to interleaved planning for long-horizon tasks where conditions change. The planner output should always be JSON or a strict schema, not freeform text, so the executor can parse it without an extra LLM call.

Add an explicit planner when the task has more than 3 to 5 steps, when steps have dependencies, or when you need a human to approve a plan before execution. For short ad-hoc tasks the LLM's implicit planning inside ReAct is sufficient.

• LangGraph plan-and-execute
• Anthropic: Building Effective Agents

Related Terms