Plan-and-Solve Agents
Measure twice, cut once. Separating strategy from execution.
1. The Chaos of ReAct
ReAct agents are "myopic." They only see the next step. If you ask ReAct to "Write a novel," it will try to write Chapter 1, then Chapter 2... and by Chapter 3 it forgets the plot. It wanders.
Plan-and-Solve separates the "Architect" from the "Builder."
2. The Planner Pattern
Step 1: The Planner A prompt that cannot use tools. Its only job is to output a checklist. User: "Research the history of coffee and write a blog post." Planner:
- Search for coffee origins (Ethiopia).
- Search for spread to Europe.
- Search for modern wave.
- Write section 1.
- Write section 2.
Step 2: The Executor A ReAct agent that takes Item #1, executes it, returns the result. Then takes Item #2.
3. Replanning
Real life breaks plans. "Search for Ethiopian records" might return 0 results. A robust Plan-and-Solve agent has a Replanner Step. If Step 1 fails, it goes back to the Architect: "I couldn't find records. Update the plan." New Plan: "Search for 'Legend of Kaldi' instead."
4. Summary
Small tasks -> ReAct. Big projects -> Plan-and-Solve. You wouldn't build a house without blueprints. Don't let your AI build software without a plan.
Key Intuition: "First we think. Then we do."