Multi-Agent Patterns: Planner-Executor

As tasks get more complex, a single agent often gets "Confused." It tries to do too many things and loses track of the goal. The Planner-Executor pattern solves this by separating the Strategy from the Action.

1. The Roles

• The Planner: A smart, slow model (e.g., Claude 3.5 Sonnet). It looks at the big goal and breaks it into a 5-step checklist.
• The Executor: A fast, specialized agent (e.g., Llama 3 8B). It takes one step at a time from the checklist and runs the tools.

2. Why this is better

1. Safety: The Planner can review the Executor's work after every step.
2. Cost: Use the expensive model only for the "Thinking" and the cheap model for the "Doing."
3. Isolation: If the Executor fails at step 3, the Planner can decide to retry or skip to step 4 without crashing the whole mission.

---

3. Visualizing the Pattern

graph TD
    User[Goal: Market Report] --> P[Planner Agent]
    P --> S1[Step 1: Get Stock]
    P --> S2[Step 2: Get News]
    P --> S3[Step 3: Graph Data]
    
    S1 --> E[Executor Agent]
    E -->|Tools| Data[Market Data]
    Data --> P
    P -->|Approved| S2

---

4. Engineering Tip: Shared State

For this to work, both agents need a Shared Memory (often a simple database or a JSON object). The Planner writes the "To-do List" to the state, and the Executor writes the "Results" to the state.

---

Summary

• Split Thinking and Doing to reduce agent confusion.
• Planner models the strategy; Executor handles the tool calls.
• This pattern enables Cost Optimization across models.
• Shared State is the glue that connects the team.