Hierarchical Planning

Generate a high-level plan first, then expand each step into a detailed sub-plan, avoiding the generate-hard versus execute-hard granularity bind.

Plan granularity is a design choice. Detailed plans are hard to generate and easier to execute; high-level plans are easier to generate and harder to execute. Hierarchical planning sidesteps the bind: produce a coarse roadmap first, then expand each step into a finer sub-plan — possibly with the same planner or a different one.

This pairs with plan-validate-execute: validate the high-level plan before expensive expansion, and validate each sub-plan before tool calls. For LLM agents, hierarchy also helps humans supply a coarse plan the agent elaborates, which is a common human approval gates pattern for complex tasks.

Generative agent architecture pushes the same idea into daily life simulation: sketch a day in broad strokes, decompose into hour chunks, then into 5–15 minute actions, store plans in the memory stream, and replan from the reaction point when perception interrupts the agenda — generate only near-horizon detail just in time when plans change often.

---

Sources

• AI Engineering (Chip Huyen), ch. 6
• Generative Agents (Park et al.), pp. 1–22