Agency-Agents: Running a Multi-Agent Team in Claude Code

8-minute read · September 2026

Every "multi-agent framework" I've tested in the last year has the same failure mode: the agents talk to each other beautifully and ship nothing. msitarzewski/agency-agents is the first one where the output actually reached my filesystem.

I ran it against CrewAI and AutoGen on the same brief: "Build a landing page for a fictional coffee brand, including copy, HTML, and a hero image prompt." Below is what I learned.

The Core Idea: A Digital Agency

Instead of abstract "agent 1 talks to agent 2", agency-agents models a small creative agency:

• Account Manager — clarifies the brief, owns scope.
• Designer — writes the design brief, chooses palette + typography.
• Engineer — writes the code.
• QA — runs the output, files issues.

Each role has its own system prompt, its own tool allowlist, and a clear hand-off protocol. It's not magic — it's a workflow dressed as agents.

Setup (5 Minutes)

You'll see the roles taking turns in the terminal. Files land in ./output/.

What It Got Right vs CrewAI and AutoGen

The differentiator isn't the LLM — it's the hand-off discipline. agency-agents forces each role to output a structured artifact (brief → spec → code → QA report). CrewAI leaves that to prompts. AutoGen leaves it to the agents themselves. Guess which one converges.

Where It Struggled

• Long tasks (10 hand-offs) still drift. Same weakness as every framework.
• Tool integration is manual — no MCP wrapper yet. I opened an issue.
• No web UI. Terminal-only. Fine for me, might not be for you.

When to Reach For This

• Content pipelines (blog + hero image + social snippet in one run).
• Landing-page prototypes.
• Any "kick off, walk away, come back to a deliverable" workflow.

For code-generation on real repos, I still prefer Claude Code with MCP servers — the agent has real filesystem access and the feedback loop is tighter.

Bottom Line

If you've bounced off CrewAI and AutoGen thinking "cool demos, useless output" — try agency-agents. It's the framework that finally made multi-agent feel shippable to me.

Star the repo, send msitarzewski a PR, and let me know what you build.