Build a Multi‑Agent Workflow – Playnex Starter Path

Create a Multi‑Agent Pipeline

Multi‑agent systems are one of the most exciting frontiers in modern AI. Instead of relying on a single model to do everything, you orchestrate multiple specialized agents — each with its own role, strengths, and perspective. This mirrors the approach used in early autonomous frameworks like AutoGPT and BabyAGI, as well as more recent research into LLM‑as‑Agents.

In this step, you’ll build a simple but powerful 3‑agent workflow:

Researcher — gathers clear, factual information
Writer — transforms research into structured prose
Editor — improves clarity, tone, and readability

Each agent is just a function that calls your local model with a different system prompt — but together, they form a coordinated pipeline capable of producing high‑quality output.

1. Create `multi.js`

touch multi.js

Add the following code:

import ollama from "ollama";

async function callAgent(systemPrompt, userPrompt) {
  const response = await ollama.chat({
    model: "llama3",
    messages: [
      { role: "system", content: systemPrompt },
      { role: "user", content: userPrompt }
    ],
    stream: false
  });

  return response.message.content;
}

// Agents
async function researcher(topic) {
  return await callAgent(
    "You are a research agent. Gather clear, factual information.",
    `Research this topic: ${topic}`
  );
}

async function writer(research) {
  return await callAgent(
    "You are a writing agent. Turn research into a well-written article.",
    research
  );
}

async function editor(draft) {
  return await callAgent(
    "You are an editing agent. Improve clarity, tone, and flow.",
    draft
  );
}

// Pipeline
async function runPipeline(topic) {
  console.log("Researching...");
  const research = await researcher(topic);

  console.log("\\nWriting...");
  const draft = await writer(research);

  console.log("\\nEditing...");
  const final = await editor(draft);

  console.log("\\n--- Final Output ---\\n");
  console.log(final);
}

runPipeline("How solar panels work");
      

This pipeline demonstrates a core principle of multi‑agent systems: specialization. Each agent focuses on a single task, and the quality of the final output improves dramatically as a result.

2. Run the Pipeline

node multi.js

You’ll see output like:

Researching...
Writing...
Editing...

--- Final Output ---
[polished article]
      

You now have a functioning multi‑agent system running entirely on your machine — no cloud, no API keys, and no external dependencies.

3. Customize Your Agents

Multi‑agent pipelines become far more powerful when you tailor each agent’s personality and role. Try modifying the system prompts:

Make the writer more creative or more technical
Make the editor more strict or more conversational
Add a “fact‑checker” agent to verify claims
Add a “summarizer” agent to condense long research

With just a few lines of code, you can build research studios, writing teams, coding assistants, analysis pipelines, and more — all powered by local models.

4. Add Tools to Agents

When you combine multi‑agent workflows with your tool system, your agents can:

read files and documents
write reports or summaries
organize folders or rename files
analyze logs or datasets
generate drafts and refine them

This is the foundation of real agentic software — systems that can reason, act, and collaborate to achieve complex goals.

Troubleshooting

Agents produce repetitive output

Strengthen or clarify the system prompts
Try a different model (Qwen and Mistral often produce more structured output)

Pipeline feels slow

Use smaller models like phi or qwen
Reduce prompt length or intermediate output

Output is inconsistent

Provide more structure in system prompts
Break tasks into smaller, more explicit steps