Task decomposition
The agent reasoning step where a high-level goal is broken into ordered, executable sub-tasks before any tool call is made. Foundational to plan-and-execute, ReAct, and tree-of-thoughts patterns.
Task decomposition is the first useful thing an agent does. It receives a vague goal ("write the launch blog post") and splits it into a plan: research the product, draft an outline, write each section, fact-check, format. Each sub-task is concrete enough to attempt with a single tool call or a short reasoning step.
Good decomposition is the single biggest predictor of agent success on long tasks. Empirically, agents that spend more time decomposing and less time acting recover from errors better and produce more useful results.
In 2026, decomposition is often handled by a dedicated planner LLM call (cheaper, faster model) separate from the executor. This separation is the basis of plan-and-execute, supervisor-worker, and hierarchical-agent patterns.
Frequently asked
Do all agents decompose tasks?+
No. Simple tool-use chatbots act on the next obvious step. Real agents — anything operating over multi-step goals — decompose explicitly. Without decomposition, the agent loses context and loops.
Should I store the decomposition or regenerate it?+
Store it. Regenerating mid-task wastes tokens and risks drift. Most production agents persist the plan to memory and update it only on explicit replan triggers.