Option 3 — Specialist Swarm

Concept landed: Skills, plugins & sub-agents Tech: Claude Managed Agents multi-agent + custom Skills + the pre-built docx skill Time: 60 minutes Output: A coordinator agent that fans work out to 3-5 specialist sub-agents, each with its own skills, that assemble a real branded Word document.

The pitch

This is the architecture that wins the next $50M transformation deal: coordinator + specialists + skills. It maps directly to how every services firm structures real work. A senior partner orchestrates; specialists (legal, pricing, technical) own their lanes; the senior partner synthesises and delivers.

You're going to build exactly that, in 60 minutes, around a Deal Desk scenario. Drop an RFP in, get a branded response doc out, watch the parallelism happen in real time on the events stream.

Setup (5 min)

You need a workspace API key on the Console (multi-agent is currently in research preview — your workspace may need to be granted access).

cd 03-specialist-swarm
pip install -r requirements.txt
export ANTHROPIC_API_KEY="sk-ant-..."

Pick a scenario card

Three cards in scenario-cards.md. Each gives you a coordinator + a different roster of specialists. Pick one. Different teams should pick different cards.

Core build (25 min)

1. Create the specialists. Run python create_specialists.py. This creates 3-4 sub-agents (Pricing, Legal, Technical Fit, Competitive) and saves their IDs to .specialist_ids.json.

2. Create the coordinator. Run python create_coordinator.py. This creates the coordinator agent with multiagent: coordinator config, listing the specialists in its callable roster.

3. Upload the skills. Run python upload_skills.py. This packages the custom skills in skills/ and uploads them via the Skills API. Each specialist gets the skill that matches its domain.

4. Run the deal. Run python run_deal_desk.py. This:

   • Uploads the synthetic RFP (synthetic-data/rfp-acme-corp.md) as a file
   • Starts a session against the coordinator
   • Asks the coordinator to produce a full proposal response
   • Streams the events so you can watch the parallel thread fan-out
   • Saves the final docx to outputs/proposal-response.docx

By minute 30 you have a Word document in outputs/, generated by a coordinator + specialists who each used their own skill.

Stretch goals (20 min)

See stretch-goals.md. The big ones:

• Custom firm-voice skill — codify your own firm's voice (every services firm has one)
• Critic sub-agent — add a fifth agent that reviews the coordinator's draft before it's finalised
• Memory across deals — coordinator remembers past wins and re-uses the right ones
• Synthetic MCP for past wins — wire up a fake CRM to the pricing specialist

Two-minute demo

Two-monitor setup:

• Monitor 1: the events stream from the coordinator session, scrolling. You'll see session.thread_created × 4, parallel running, then agent.thread_message_received flowing back. The visible parallelism IS the demo.
• Monitor 2: open outputs/proposal-response.docx. Real document, branded, ready to send.

Narrate the events stream while it runs. The room will get it.

What's in this folder

03-specialist-swarm/
├── README.md
├── scenario-cards.md
├── stretch-goals.md
├── requirements.txt
├── create_specialists.py          (creates the sub-agents)
├── create_coordinator.py          (creates the coordinator)
├── upload_skills.py               (uploads custom skills via Skills API)
├── run_deal_desk.py               (runs the full swarm against an RFP)
├── stretch_critic_subagent.py     (stretch: critic agent)
├── skills/                        (custom skills, one per specialist)
│   ├── pricing-playbook/SKILL.md
│   ├── legal-checklist/SKILL.md
│   └── competitive-intel/SKILL.md
└── synthetic-data/
    ├── rfp-acme-corp.md           (the RFP that triggers the swarm)
    ├── past-wins.json             (used by pricing specialist)
    └── product-overview.md        (used by technical specialist)