Planning

Process agents published

Also known as: Task Planning, Agent Planning, Plan Generation

Definition

The process by which an agent formulates a sequence of steps or strategy to achieve a goal before executing actions. Planning separates "what to do" from "doing it," allowing the agent to reason about the approach, identify potential obstacles, and optimize the path. Planning can happen once upfront or be interleaved with execution (replan as you learn).

What this is NOT

  • Not the same as reasoning (planning is reasoning about future actions, not all reasoning is planning)
  • Not a workflow definition (plans are generated by the agent, workflows are predefined)
  • Not Chain-of-Thought (CoT is step-by-step reasoning within a single response; planning structures multiple actions)

Alternative Interpretations

Different communities use this term differently:

classical-ai

Automated generation of action sequences to transition from an initial state to a goal state, given a model of actions and their effects. Studied in STRIPS, PDDL, and hierarchical task network (HTN) planning.

Sources: Ghallab et al.: Automated Planning (2004), STRIPS and PDDL literature

llm-practitioners

Prompting an LLM to decompose a complex task into subtasks before execution, often producing a list of steps that can be executed sequentially or adapted during execution.

Sources: Plan-and-Solve paper (2023), HuggingGPT paper (2023), Least-to-Most prompting

Examples

  • An agent decomposes 'write a blog post' into: research topic, create outline, draft sections, edit, publish
  • A coding agent plans: understand spec, design architecture, implement modules, write tests, integrate
  • A travel agent plans: find flights, find hotels, check conflicts, book in order

Counterexamples

Things that might seem like Planning but are not:

  • Chain-of-Thought reasoning within a single response (that's not planning for future actions)
  • A workflow hardcoded by the developer (not generated by the agent)
  • Reactive agents that decide each action based only on current state

Relations

  • overlapsWith reasoning (Planning is a form of reasoning about actions)
  • requires agent (Agents use planning to structure their behavior)
  • overlapsWith agent-loop (Planning may occur within or before the agent loop)
  • overlapsWith reflection (Reflection can inform replanning)

See Also