Planner-executor architecture

An agent design pattern that represents proposed steps separately from the component that executes tools and returns observations.

The planner looks at the goal, breaks it into steps, and produces a plan. The executor takes each step and performs the action: calling a tool, querying a database, or writing a file. After each action, the result goes back to the planner, which can revise the remaining steps based on what happened. A research agent's planner might specify the question to investigate, which claims need evidence, and the sources to check, while the executor handles search calls and file reads. The planner and executor can be separate models, separate prompts within one model, or a mix of model and code.

Builder example

An explicit plan artifact can help diagnose whether a failure began in task decomposition or tool execution, and it can support policy checks before expensive steps. The split also adds state, latency, and synchronization failure modes. Some tasks work better with short plan-act-observe cycles than with one large plan fixed in advance.

Common confusion: The planner is no more inherently reliable than the executor. A planner with poor context produces bad plans, and a good plan still fails if the executor lacks proper error handling. Both layers need attention.