Task Decomposition

Task decomposition is the practice of breaking a complex problem into smaller, manageable subtasks that can be handled individually. In AI, this typically means splitting a difficult request into simpler steps that the model processes sequentially, producing better results than attempting everything in a single prompt.

Why decomposition improves AI results

AI models perform better on focused, specific tasks than on broad, complex ones. Asking a model to "Research, analyse, write, and format a competitive analysis report" in a single prompt overwhelms it. Breaking this into distinct steps — research competitors, analyse strengths and weaknesses, write the narrative, format the report — produces dramatically better results at each stage.

Common decomposition patterns

• Sequential: Each subtask feeds into the next. Research → Analysis → Writing → Review.
• Parallel: Independent subtasks run simultaneously. Research competitor A, competitor B, and competitor C at the same time.
• Hierarchical: A planning step identifies subtasks, which may themselves be decomposed further.
• Iterative: Repeat a subtask with refinement. Draft → Review → Revise → Review → Finalise.

Decomposition in practice

When you use chain-of-thought prompting ("Think step by step"), you are asking the model to decompose internally. When you break a project into multiple prompt calls, you are decomposing externally. External decomposition gives you more control:

• You can inspect intermediate results and correct errors before they propagate.
• You can use different models for different subtasks (a cheap model for extraction, a capable model for analysis).
• You can inject additional context or data between steps.
• You can parallelise independent subtasks for speed.

Task decomposition in AI agents

AI agents use task decomposition as a core capability. When given a complex goal, the agent:

1. Creates a plan by breaking the goal into subtasks.
2. Executes each subtask, potentially using different tools.
3. Evaluates intermediate results and adjusts the plan as needed.
4. Combines subtask results into the final output.

Effective decomposition strategies

• Start with the output: What does the final deliverable look like? Work backwards to identify the steps needed.
• Identify dependencies: Which steps must happen in order? Which can be parallelised?
• Size subtasks appropriately: Each subtask should be focused enough for the AI to handle well but not so granular that overhead dominates.
• Include quality checks: Add verification steps between subtasks to catch errors early.