Multi-agent setup

Questions
1

Describe the problem/error/question

Hi, I want to ask about multi-agent setup in n8n. If I look tutorial or a blog post many says that multi-agent is breaking down task into smaller pieces but the task that the example have is always focus on calling tool node. Now I have a different setup where I don’t want call any tool but let’s say I want to create a brainstorm multi-agent setup where one agent as the Parent is an orchestrator and the one who receive the idea from user and there’s child agent that create question another agent answering the questions and other validating the answer, does this kind of setup possible?

additional information the child agent should be an AI agent tool node.

2

@ezraluandre

Your proposed setup is entirely possible in n8n using the specialized AI Agent Tool node, which shifts the paradigm from an agent using “functional tools” (like a calculator) to an agent managing “specialist agents.” This allows you to build a hierarchical, multi-tiered structure where a primary orchestrator delegates complex cognitive tasks to sub-agents rather than just executing simple data-fetching commands.

In your brainstorming example, you would design a Parent Agent that serves as the orchestrator, equipped with three distinct AI Agent Tool nodes: a Questioner, a Responder, and a Validator. The Parent receives the user’s idea and manages the conversational flow, passing the context sequentially through these specialists—generating probing questions, providing detailed answers, and finally validating the quality of the result—all within a single, cohesive workflow.

This architectural approach is highly effective because it replaces a single, over-complicated prompt with several small, highly focused system prompts. By breaking the logic into specialized roles, you significantly increase the accuracy of the results, reduce the likelihood of hallucinations, and make the entire multi-step process much easier to debug and manage from a single canvas.

3

Something like this?

4

Hi @ezraluandre, yes. The parent calls each child as a tool, but the children don’t need external tools for this text-only task. The JSON above is a different setup: two regular AI Agent nodes in sequence, plus Date/Time and MCP tools.

Connect Questioner, Answerer and Validator as three AI Agent Tool nodes to the parent AI Agent’s Tool connector. Give each a Chat Model and a clear Description. In the parent’s System Message, tell it to pass the idea to Questioner, its questions to Answerer, then both questions and answers to Validator. Use $fromAI() in each child’s Prompt (User Message) to receive that input.

The parent chooses which child to call, so the prompt cannot guarantee that order. If every stage must run in sequence, use three Basic LLM Chain nodes on the main workflow path instead.

6

Thanks for answering my question, the next question that I want to ask is how to make the parent agent do the orchestration?
In the example that I found people usually stating like “if you need to send email use Email agent” or “if you need to do research use research agent” but I don’t think I could stated that “if you need to create question, you should call question agent” I think that is kind of off because I always need to do that and doing so could make parent agent think that “I don’t think I need to call the question agent so I won’t call it”, how to deal with this kind of issue?

7

not really, it is more like in this video https://youtu.be/lW5xEm7iSXk?si=Mwhgn-HwIkEw8CAo

8

Hi, thanks for answering my question. I have a question about what you stated, you said

If every stage must run in sequence, use three Basic LLM Chain nodes on the main workflow path instead.

Does that means if every stage must run in sequence rather than using AI agent tool node it is better to use AI Agent node but in sequence, is this the correct interpretation?

9

AI Agents are non-deterministic. If you want to avoid hallucination, state everything explicitly, like

As for

@Anshul_Namdev has already answered your question

10

@ezraluandre

Yes, this works, and tool calling is still how it works. In n8n, a sub-agent is a tool. The parent decides when to call each child, just as it would call an HTTP or Sheets tool.

Setup:

  1. Chat Trigger → AI Agent (Orchestrator). This agent receives the user’s idea.
  2. Attach three AI Agent Tool nodes to the orchestrator’s Tool input, each with its own model and system prompt:
    • Question Generator: “Given an idea, produce 5 probing questions about feasibility, audience, risks…”
    • Answerer: “Answer each question concretely using the idea’s context.”
    • Validator: “Check each answer for gaps, contradictions or weak reasoning. Return PASS or FAIL with reasons.”
  3. The orchestrator’s system prompt sets the order: “Call Question Generator with the idea. Pass its questions plus the idea to Answerer. Pass the Q&A to Validator. If FAIL, send the failed items back to Answerer (max 2 retries). Then summarize the final brainstorm for the user.”

Things that trip people up:

  • Tool descriptions matter a lot. The orchestrator picks a child based on its description, so write it clearly: what the agent does, what input it expects, and what it returns.
  • Child agents don’t share memory. The orchestrator has to pass the full context (the idea plus prior outputs) in every call.
  • Set Max Iterations on the orchestrator so a validate→retry loop can’t run forever.
  • Use a cheaper model for the children and a stronger one for the orchestrator.

An alternative: if the order is always the same (generate → answer → validate), you don’t need an orchestrator. Chain three plain AI Agent nodes and use an IF node after the Validator to loop back. This is cheaper, predictable and easier to debug. The orchestrator approach is worth it when the parent should decide dynamically, for example skipping validation for simple ideas.

PRO TIP- building this with Claude/Codex via the n8n MCP and watching what it creates is a fast way to learn."

11

One cost gotcha with the parent-calls-children-as-tools pattern: every child call usually re-sends the orchestrator’s growing context plus the child’s own system prompt. If the parent calls Questioner, then Answerer, then Validator on every user idea, that’s 3 model round-trips and often 3x the input tokens of a straight sequence of chains.

If the stages always run in that order (your brainstorm case), I’d use three Basic LLM Chain or AI Agent nodes on the main path instead of tool-calling. You still pay for three calls, but each prompt stays small and the order is deterministic, with no orchestrator deciding to skip or double-call a specialist. Keep tool-style sub-agents for optional branches (research only sometimes, email only sometimes).

Also log usage per node (model, input, output) even for text-only agents. A validator that echoes the full Q&A back into its prompt is a quiet token multiplier and the first place I’d trim if a run feels expensive.

12

Hey @ezraluandre

I’ll audit your n8n agents end-to-end, break them in staging, and hand them back client-ready with every failure point fixed and documented.

Here’s the approach:

Full workflow review across n8n, Vapi, Twilio, ElevenLabs, OpenAI/Claude

Edge-case + error-handling stress tests (timeouts, retries, webhook failures, bad inputs)

Integration/API audit — auth, rate limits, payload mismatches, flaky nodes

AI behaviour QA — prompt drift, hallucinations, tool-call reliability, voice quality

Fixes shipped in-place with a short bug log + hardening recommendations per agent