Planning & Reasoning

Introduction

Planning turns a high-level goal into a sequence of doable steps. Reasoning checks each step makes sense. Strong planners are the difference between an agent that finishes the task and one that loops.

Beginner analogy: trip planning — pick city, dates, flight, hotel, activities. An agent does the same for any goal.

Understanding the topic

Core concepts:

• Plan-then-execute: one big plan upfront.
• ReAct: plan one step at a time, observe, replan.
• Tree-of-Thought: explore branches, pick best.
• Re-planning on failure is essential.
• Use a smaller / cheaper LLM for execution, big one for planning.

Syntax reference

Visual workflow / architecture:

Goal: "Find a flight under $400 to Paris next week"
    │
    ▼
Plan
  1. Search flights
  2. Filter by price < $400
  3. Pick earliest departure
  4. Return summary
    │
    ▼
Executor runs each step
    │
    ▼
Failure? → replan from current state

Real-world use

Devin plans whole PRs; OpenAI Deep Research plans multi-source investigations; CrewAI uses hierarchical plans.

Best practices

• Always allow re-planning — the world changes mid-run.
• Show the plan in the trace UI — humans debug 10x faster.

Common mistakes

• Rigid upfront plans that can't adapt to tool failures.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Three planning strategies.
• Q2. When is ReAct better than plan-then-execute?
• Q3. Scenario: your agent always tries the same broken plan. What's wrong?