Planner (agents)
The agent or component that turns a goal into an explicit plan of steps, before or during execution, for itself or for other agents to carry out.
A planner is the part of an agent system that turns a goal into an explicit sequence of steps, either as a separate agent that writes the plan or as a planning phase built into an agent.
Forms it takes.
- Inside the model’s turn. Google ADK lets you attach a planner to an
LlmAgent.BuiltInPlanneruses the model’s own thinking feature, with a budget for thinking tokens.PlanReActPlannertells models without that feature to write a plan first, then act, for example by calling tools, and give reasoning for each step. - A planning pass before execution. Setting
planning=Trueon a CrewAI crew runs an AgentPlanner before each crew iteration. It plans the tasks step by step and adds that plan to each task’s description, andplanning_llmchooses the model that plans. - A lead agent that plans and delegates. In Anthropic’s research system, the lead agent works out a strategy, saves the plan to memory, and spawns sub-agents to carry it out. It saves the plan because a long run can exceed the context window and truncate it.
- Built into a harness. Microsoft Agent Framework’s Harness Agent includes planning and to-do tracking for long, multi-step tasks.
Why plan explicitly. A written plan gives the agent something to check progress against, lets a person review the approach before anything runs, and survives context loss when it is stored outside the context window. The cost is that a plan made before any tool results can go stale, so it has to be revisable as results arrive.
Plans between agents. Across organizations, plans stay private. A2A lets agents collaborate on declared capabilities without sharing internal plans, so a planner can decide which remote agent to ask for what, but it sees only the task states and artifacts that come back.
Neighbouring terms. An orchestrator agent often plans and delegates in one role. Agent memory is where long plans are kept.