C2 · Plan and execute
Family: Single agent
A planner writes the whole plan, executors carry it out, a solver writes the answer.
Principle
A planner first produces the complete plan, sometimes as a dependency graph. Executors, tools or small models, carry it out without rereading the whole context, then a solver synthesises. The large model steps in only to plan, replan and conclude.
The problem it solves
In an agent loop, every step rereads the whole history: cost and latency climb fast, and the plan exists nowhere you can read it.
When to use it
- The tool pipeline is fairly predictable: aggregating several sources, independent queries to run in parallel.
- Budget or latency is tight, or you want a readable, auditable plan before anything runs.
When to avoid it
- Each step truly depends on the result of the previous one, as in exploration: prefer the agent loop (C1).
- The plan is neither validated nor replanned along the way: a wrong plan makes the whole run fail.
Synonyms by vendor
- Plan-and-Solve (Wang et al. (ACL 2023))
- ReWOO (Xu et al. (2023))
- LLMCompiler (Kim et al. (ICML 2024))
- Planning (A. Ng, A. Gulli (ch. 6))
- Single-path / multi-path plan generator (CSIRO Data61)
Neighbouring patterns
Sources
- ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models — B. Xu et al., 2023
- An LLM Compiler for Parallel Function Calling — S. Kim et al. (ICML 2024), 2024
- Agentic Design Patterns, part 4: Planning — DeepLearning.AI, The Batch (A. Ng), 2024-04-10
- Agents — Chip Huyen, 2025-01-07
Content reviewed on 2026-10-02