Building Your First Multi-Agent System with Marblo — Hands-On Tutorial

What You'll Build

By the end of this thirty-minute tutorial, you'll have a working multi-agent workflow that:

1. Takes a topic as input
2. Runs a researcher agent on Claude that gathers facts via MCP web search
3. Passes findings to a writer agent on GPT-4.1 that drafts a 600-word article
4. Routes the draft to a fact-checker agent on Gemini for independent verification
5. Returns the final article with citations

Everything in this tutorial assumes Marblo Pro (or higher) — the free tier supports single-agent workflows but not the multi-agent orchestration we use here.

Prerequisites

• Marblo account (Free is fine for the first agent, Pro+ for the full multi-agent flow)
• API keys for Anthropic Claude, OpenAI GPT, and Google Gemini
• 30 minutes

Step 1 — Create the Board

A "board" in Marblo is one workflow. Each board has agents, MCP tools, and a routing graph.

1. Open Marblo → New Board
2. Name it: article-pipeline-tutorial
3. Description: Three-agent workflow: research, write, verify

That's the setup. Total elapsed time: 90 seconds.

Step 2 — Add the Researcher Agent

The researcher's job is to gather facts. We use Claude because long-context reasoning on web results outperforms generation-optimized models.

1. Click Add Agent
2. Role: researcher
3. Model: Claude Opus 4.7
4. System prompt:

You are a research agent. Given a topic, search the web for the most
authoritative sources and produce structured findings.

Output format:
- Topic summary (2-3 sentences)
- 5-8 key claims, each with a citation URL
- 2-3 areas of disagreement or uncertainty in the sources

Do not generate prose. Output only structured findings.

5. Add MCP Tool: web_search (Marblo's built-in web search tool)
6. Save Agent

Step 3 — Add the Writer Agent

The writer turns structured findings into an article. GPT-4.1 wins on speed and fluency for high-volume generation.

1. Click Add Agent
2. Role: writer
3. Model: GPT-4.1
4. System prompt:

You are a writing agent. You receive structured research findings and produce
a 600-word article in a confident, direct voice.

Rules:
- Use only facts from the provided findings
- Cite sources inline as [Source: URL]
- Lead with the answer, not background
- One short paragraph per idea
- No hedging language ("might," "could," "perhaps") unless the source is uncertain

5. Input: receives output from researcher
6. Save Agent

Step 4 — Add the Fact-Checker Agent

The fact-checker verifies the article against the original findings. Critically, we use a different vendor (Gemini) to avoid the verifier inheriting the writer's blind spots.

1. Click Add Agent
2. Role: fact_checker
3. Model: Gemini Pro
4. System prompt:

You are a fact-checking agent. You receive:
1. Original research findings (structured)
2. A draft article

For each factual claim in the draft, verify it appears in the findings.
Output:
- "VERIFIED" claims (supported by findings)
- "UNSUPPORTED" claims (not in findings — flag for removal)
- "CITATION_MISSING" claims (in findings but missing inline citation)

Be strict. If a number, date, or attribution doesn't exactly match the findings,
flag it.

5. Inputs: receives output from researcher AND writer
6. Save Agent

Step 5 — Connect the Graph

In Marblo's graph view, drag connections:

[input] → researcher → writer → fact_checker → [output]
                    ↘____________↗
              (researcher findings also flow to fact_checker)

The second connection (researcher → fact_checker) is what makes the verification independent. The fact-checker sees both the original findings and the writer's draft, and can flag discrepancies.

Step 6 — Test the Board

1. Click Run Test
2. Input: "The state of AI agent orchestration in 2026"
3. Watch the trace view as each agent runs

Expected output (abbreviated):

researcher: 7 findings gathered, 4 citation URLs
writer: 612-word draft produced
fact_checker:
  - 7 VERIFIED claims
  - 0 UNSUPPORTED
  - 1 CITATION_MISSING (the writer dropped a citation on claim #4)

The fact-checker caught a missing citation. That's the value of the heterogeneous setup — if writer and verifier were the same model, the missing citation would have slipped through.

Step 7 — Deploy

1. Click Deploy
2. Marblo gives you a webhook URL
3. POST { "topic": "your topic here" } to that URL
4. Receive { "article": "...", "trace_id": "..." }

That's it. Production endpoint. Versioned, traced, cost-attributed per agent.

What Just Happened

You ran a workflow that:

• Used three different vendors (Anthropic, OpenAI, Google) — heterogeneous by design
• Each agent ran on the model that fits its role (reasoning, generation, verification)
• The fact-checker was independent from the writer (different vendor)
• Cost was attributed per agent (Marblo logs this automatically)
• You can trace any production request end-to-end

Total cost for a single run on our test: $0.038. Single-model equivalent would have been around $0.09. Multiply by 10K runs/month and the savings pay for the Pro subscription several times over.

Common Beginner Mistakes

Mistake 1 — Using the same model for all three agents The cost-quality tradeoff disappears. Use heterogeneous from day one.

Mistake 2 — Skipping the second connection (researcher → fact_checker) Without it, the fact-checker has no ground truth. It can only check internal consistency, not factual accuracy.

Mistake 3 — Vague system prompts The structured output format in the researcher prompt is what lets the writer parse it cleanly. Specificity in prompts pays off across the chain.

Mistake 4 — Skipping the test run Marblo's test view shows the trace. Running it once before deploy catches 80% of the bugs.

What to Build Next

This three-agent pattern is the foundation. Common extensions:

• Add an editor agent as a final pass for tone and style
• Add a localizer agent for multi-language output
• Add an MCP tool to write the final article to your CMS
• Branch on fact-checker output — if UNSUPPORTED claims exist, loop back to writer

Each extension is one more agent in the graph. The pattern scales.

When You Need Help

If you get stuck or want a 30-minute walkthrough with a Marblo team member on your specific workflow, book a session. We do this regularly for teams onboarding to Marblo.

Further Reading

• AI Agent Orchestration Platforms in 2026 — A Comparison
• Heterogeneous Agents in Production
• MCP Servers in Production

---

Last updated: 2026-05-19. Marblo evolves quickly — UI labels may shift slightly, but the workflow pattern is stable.