Task Decomposition

Introduction

Task decomposition is splitting a goal into smaller, doable subtasks. Good decomposition is the most underrated agent skill — it determines whether the run is 5 steps or 50.

Beginner analogy: moving house — 'move' is overwhelming; pack→hire→drive→unpack is doable.

Understanding the topic

Core concepts:

• Prompt the model: 'break this into subtasks'.
• Limit subtask count (3-7) to avoid over-fragmentation.
• Each subtask should map to ONE tool call or LLM call.
• Use a planner LLM, even if executor is cheap.
• Allow merging or skipping subtasks dynamically.

Syntax reference

Visual workflow / architecture:

Goal: "Launch a Twitter campaign for product X"
     │
     ▼
Decompose:
  1. Draft 5 tweet ideas (LLM)
  2. Pick best 3 (critic)
  3. Generate images (tool)
  4. Schedule with Buffer API (tool)
  5. Report back (LLM)

Real-world use

Devin decomposes PRs into file edits; Operator decomposes orders into clicks; OpenAI Deep Research into questions.

Best practices

• Bound subtask count.
• Test the planner alone — bad plans poison every run.

Common mistakes

• Over-decomposing into 30 micro-steps — context explodes.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why limit subtask count?
• Q2. How do you test a planner in isolation?
• Q3. Scenario: your agent over-decomposes. Two fixes?