task-decomposition
Task decomposition — the planning half of the agent anatomy: breaking a large, multi-step task into smaller, manageable subgoals so the agent can handle complexity (Weng 2023). “A complicated task usually involves many steps. An agent needs to know what they are and plan ahead.”
Four ways to decompose
Per the source:
- LLM with simple prompting —
"Steps for XYZ.\n1."or"What are the subgoals for achieving XYZ?". Cheapest; the model proposes its own plan. - Task-specific instructions — e.g.
"Write a story outline."for novel-writing. The decomposition structure is supplied by the task, not invented. - Human inputs — a person supplies the breakdown.
- External planner — see LLM+P below.
The reasoning-based methods
- Chain of Thought (Wei et al. 2022) — “think step by step” is a decomposition: it turns one hard task into a sequence of simpler ones.
- Tree of Thoughts (Yao et al. 2023) — decompose into thought steps, branch into multiple candidates per step, search the tree (BFS/DFS) with state evaluation. Decomposition + exploration. (Detailed on chain-of-thought.)
- LLM+P (Liu et al. 2023) — a distinct, outsourced approach: the LLM translates the problem into PDDL (Planning Domain Definition Language), hands it to a classical planner for the actual long-horizon planning, then translates the PDDL plan back into natural language. The planning is done by a symbolic tool, not the LLM — works where domain-specific PDDL + a planner exist (common in robotics, rare elsewhere).
Where this shows up in the 2026 wiki
- plan-mode — a read-only mode that produces and refines a decomposition before any file is touched; the engineered descendant of “plan ahead”.
- long-running-agents — long-horizon tasks are split (initializer + coding agent), with subgoal state externalized to files. The source’s own conclusion flags long-term planning and task decomposition as a core limitation: LLMs “struggle to adjust plans when faced with unexpected errors.”
- multi-agent-orchestration — an orchestrator decomposes and delegates subgoals to workers; HuggingGPT’s task planning stage (tool-use) is an early instance.
- Pairs with self-reflection — decomposition sets the plan; reflection corrects it when steps fail. Together they are Weng’s “Planning” component.
Cross-link
- 2026-06-04-llm-powered-autonomous-agents — the source.
- chain-of-thought — CoT/ToT, the reasoning-based decomposition methods.
- self-reflection — the other half of Planning.
- llm-agent — the anatomy this sits inside.
- plan-mode, long-running-agents — engineered descendants.