Plan-and-Execute

Plan-and-execute is the architectural pattern that separates planning from acting. In phase one, the planner LLM produces a structured plan (a list of steps with dependencies). In phase two, the executor carries out each step, either with a separate model call per step or by handing each to a specialist agent. The pattern is useful when the work is non-trivial and benefits from auditability before action begins. LangGraph documents this as the canonical pattern for long-horizon tasks. The opposite pattern is interleaved reasoning (ReAct), where the agent decides each step right before executing it.

The tradeoff: full upfront plans give you a checkpoint where a human can review before any tools are called, but they go stale if early steps change the world. Interleaved reasoning (ReAct) is more adaptive but harder to audit before execution. The hybrid pattern that wins in practice: generate a plan upfront, execute step-by-step, and re-plan when an executed step's observation contradicts the plan's assumptions. This combines auditability with adaptability. LangGraph's "plan-and-execute" tutorial implements this exact pattern.

Use plan-and-execute when the task has 5+ steps, when steps have dependencies, when human approval of the plan is valuable, or when long-horizon work needs an audit trail.

• LangGraph: Plan and execute tutorial
• Anthropic: Building Effective Agents

Related Terms