Course contentsShow
← Machine Learning and Deep Learning
Lesson 2092 of 3,53844. AI Agents: FoundationsFree lesson

Tree-of-Thoughts for Agent Planning

Exploring multiple planning branches in parallel and selecting the most promising path forward.

Tree-of-Thoughts for Agent Planning

What you'll learn: How agents can explore multiple planning paths simultaneously and pick the best strategy by evaluating competing thought branches.

The Core Idea

While Chain-of-Thought (lesson 2088) generates a single linear sequence of reasoning steps, Tree-of-Thoughts (ToT) explores multiple reasoning paths in parallel, like branches on a tree. The agent generates several possible next steps, evaluates each one, and either continues the most promising branches or backtracks to try alternatives.

Think of it like chess: a strong player doesn't just consider one move sequence—they explore multiple possible futures ("If I move here, they might respond like this... or like that..."), evaluate which paths look most promising, and choose accordingly.

How It Works

  1. Branch Generation: At each decision point, the LLM generates multiple candidate "thoughts" or sub-plans (e.g., 3-5 different approaches).
  2. Evaluation: Each branch is scored or evaluated—either by the LLM itself ("rate this plan 1-10") or through simulation/feedback.
  3. Selection: Keep the top-k most promising branches and prune weak ones.
  4. Expansion: Continue developing the selected branches deeper, generating next-level thoughts.
  5. Iteration: Repeat until a solution is found or a depth limit is reached.

This differs from ReAct (lesson 2087) or Plan-and-Execute (lesson 2089), which typically follow a single trajectory with occasional replanning. ToT systematically explores the planning space, not just a single path through it.

When to Use ToT

Tree-of-Thoughts shines for:

  • Complex reasoning tasks with multiple valid approaches
  • Problems where early decisions significantly impact outcomes
  • Situations requiring strategic lookahead
  • Tasks where you can afford the computational cost of exploring multiple branches

Key Takeaway: Tree-of-Thoughts transforms agent planning from a linear process into a search problem, exploring multiple reasoning branches simultaneously and selecting the most promising path through explicit evaluation and pruning.