Task Decomposition

Task decomposition is the act of splitting a complex goal into atomic units. "Onboard a new customer" decomposes into "create CRM record, send welcome email, schedule kickoff call, generate API key, post to Slack." Each sub-task is small enough to assign to a specific tool or sub-agent, easy enough to verify completed, and independent enough that failures in one do not corrupt the others. Decomposition is the precursor to planning: without good decomposition, planning produces vague high-level steps that the executor cannot reliably carry out.

The art of decomposition is finding the right granularity. Too coarse and steps are still too vague to execute reliably. Too fine and the plan explodes into hundreds of trivial steps that overwhelm the model context. The rule of thumb: each sub-task should map to one tool call or one focused LLM invocation. If a sub-task needs three tool calls in sequence, split it. If it needs zero, fold it into an adjacent step. Good decomposition is half the battle in any multi-step agent system.

Required for any agent task that takes more than one tool call. Make decomposition explicit in your planner prompt (output as JSON list) so the executor can iterate cleanly.

• Anthropic: Building Effective Agents
• LangGraph: Plan and execute tutorial

Related Terms