Planning patterns

ReAct, plan-and-execute, reflection, tree of thoughts, and when modern reasoning models replace the patterns entirely.

The agent loop calls the LLM. But what does the LLM produce on each turn — a single action? A plan? An internal monologue? The patterns that shape this are called planning patterns, and the dominant ones in 2026 are different from what was dominant in 2023.

This module is being expanded with worked code for each pattern.

Coming in the next revision:

• ReAct (Reason + Act). The default — model alternates between thought and action. Why it’s still the strongest baseline.
• Plan-and-Execute. Generate a multi-step plan first, then execute each step. When it beats ReAct, when it doesn’t.
• Reflection. After each action, the agent critiques its own output. Quality wins; cost matters.
• Tree of Thoughts. Branch-and-explore for hard problems. Expensive; reserved for specific shapes.
• Self-consistency. Sample N solutions, vote. The bias-variance trade-off for agents.
• Reasoning models (o-style). When the model does its own internal chain of thought, what happens to the planning framework? (Short answer: it gets smaller.)
• Critic-actor. Multi-agent dressed as a planning pattern.

Next: Module 06 — Frameworks.