Why it matters

Complex tasks benefit from explicit planning. Understanding when planning helps vs hurts shapes agent design.

Advertisement

The architecture

Two-phase: planner LLM produces plan (list of subtasks). Executor(s) run each subtask.

Optional replanning: if execution fails, generate new plan.

Planning agent structureGoaluser requestPlanner LLMdecompose to subtasksExecutorrun eachReplanning on failure improves reliability but adds latency
Planning phase.
Advertisement

How it works end to end

Structured plans: JSON list of subtasks. Enables programmatic validation.

Dependencies: some subtasks depend on others. Plan can encode DAG.

Parallel execution: independent subtasks run concurrently.