Multi-agent pipeline: planner, executor, reviewer

One giant prompt doing everything is fragile. Three narrow agents in a row are debuggable, cheaper to iterate, and each can use a different model.

The pattern

Input → Planner (LLM) → Executor (LLM ⇄ tools) → Reviewer (LLM) → Output
                                                      └─ rejected → back to Executor / Human in the Loop

Step by step

1. Planner — LLM Agent (AI) — objective: "Break the request into an ordered list of concrete steps with the data each needs. Output JSON only." Use a reasoning-strong model here.
2. Executor — LLM Agent — receives {{planner.steps}}; objective: "Execute each step using your tools. Report what you did and the results per step." Wire its tools into this node: MCP integrations, Hive App tools.
3. Reviewer — LLM Agent — objective: "Check the execution against the plan: completeness, correctness, tone. APPROVE or REJECT with reasons." A cheaper model usually suffices.
4. Conditional Flow (Base) — APPROVE → output; REJECT → loop back to the Executor with the reviewer's notes, or escalate to Human in the Loop.

Each agent's Console shows its reasoning — three small consoles beat one giant opaque prompt.

Variations

• Isolate each role in its own flow and compose them with Sub-flow nodes for reuse.
• Add a Memory node (Data) so the Executor accumulates results across loop iterations.
• Theory behind the roles: What is an AI agent?

Genius prompt: "Create a multi-agent pipeline: a planner agent, an executor agent with my CRM as a tool, and a reviewer agent that approves or sends back with notes."