liveBook · Manning

Planning

As explained in Build a Multi-Agent System (from Scratch)

Planning is the capability of formulating and revising a sequence of actions or steps toward completing a task.

Planning is an LLM agent’s ability to formulate and revise steps toward a task. It shapes which actions the agent takes, including tool calls, and helps it respond to results as execution unfolds. section 1.2.1

A task-processing loop can use planning to:

  • Form an initial plan from the user’s request. section 1.2.1
  • Synthesize results from previous actions and judge their contribution to the task. section 1.2.1
  • Adapt the plan, course-correct, or finish early when the task is complete. section 1.2.1

A poor initial plan can produce incorrect results, timeouts, or inefficiency. Planning throughout execution gives the agent a chance to adjust when earlier actions do not lead toward the task. section 1.2.1

"I need to find all of the bakeries in New York City that sell croissants and check their prices as well as their ratings. Then, I need to build an analysis with this information to determine the best-value croissants in NYC."

For this task, an agent could plan to find bakeries, gather croissant prices and ratings, and compare the results to identify the best value. section 1.2.1

How Build a Multi-Agent System (from Scratch) covers it

How deep it goes
Explains initial planning and revision from action results, without detailing planning algorithms.
Examples
Uses a New York City croissant search requiring bakery, price, and rating comparisons.
What it assumes
Assumes a backbone LLM that can formulate plans and synthesize results from previous actions.
What you can do after
Evaluate when an agent should revise its plan or end execution based on action results.
How it connects
Connects planning with LLM agents, tool calling, and task-processing loops.

Also explained in

Compare how 11 books explain planning →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build a Multi-Agent System (from Scratch). The section links point to the parts of the book it draws on.