Workflow Topology
Pattern 09 of 9
Plan-and-Execute
Explicit planning phase produces an inspectable plan; execution phase runs it with optional replanning.
Plan-and-execute separates a task into two explicit phases. In the planning phase, a model generates a step-by-step plan for achieving the goal. The plan is a first-class artifact: human-readable, inspectable, and modifiable before execution begins. In the execution phase, an executor works through the plan, calling tools and tracking progress. Crucially, the executor can trigger replanning if it encounters a situation the original plan did not account for. The plan is the starting point, not a contract.
Why it matters
Fully autonomous agent loops are hard to observe and hard to trust. When an agent takes twenty actions without producing an inspectable intermediate artifact, it is difficult to know whether it is on the right track until it finishes. Plan-and-execute makes the agent's strategy visible before execution begins. That visibility enables human review, automated validation, and meaningful intervention if the plan is obviously wrong. It also tends to produce more coherent execution because the planning model can reason about the full task at once rather than step by step.
Deep Dive
The Plan-and-Solve paper from Wang et al., published in 2023, showed that explicitly generating a plan before execution improved reasoning quality on arithmetic and logical reasoning benchmarks compared to standard chain of thought. The improvement came from two sources: the planning step forced the model to decompose the problem before attempting to solve it, and the explicit plan reduced the chance of the model losing track of what it was doing mid-execution. These gains generalize beyond arithmetic to any task where the model can benefit from decomposing the problem before executing.
Replanning is what distinguishes plan-and-execute from ReWOO. When the executor discovers that a step is impossible, unnecessary, or producing unexpected results, it can call the planner again with the current state. The planner revises the remaining steps of the plan based on what has been learned so far. This is more expensive than pure ReWOO because replanning involves another full LLM call, but it handles the case where the initial plan was based on incomplete information. LangGraph's plan-and-execute documentation shows this pattern implemented as a graph where the executor node can route back to the planner node on certain conditions.
The inspection affordance is underrated. Most teams building agents focus on whether the agent succeeds. The plan-and-execute pattern gives you a natural point to check whether the agent is going to succeed before it spends all its tool budget on the wrong path. An automated plan validator, a simple LLM call that reads the plan and checks whether it makes sense given the goal and available tools, can catch obvious plan failures cheaply. In regulated industries or high-stakes deployments, a human checkpoint at the plan approval stage is a reasonable requirement. The plan is the right abstraction level for that checkpoint: specific enough to evaluate, abstract enough to read quickly.