A coordinator agent that fans an RFP out to a team of specialist sub-agents — and gets back a finished, branded proposal
One senior partner orchestrates. Specialists own their lanes. The work comes back assembled.
The idea • Architecture • Quick Start • Scenarios • Stretch goals
This is the architecture behind how real services firms run big deals: a coordinator + specialists + skills. A senior partner orchestrates; specialists (pricing, legal, technical, competitive) own their lanes; the partner synthesises everything into one deliverable.
Specialist Swarm builds exactly that around a Deal Desk scenario. Drop an RFP in, and a coordinator agent delegates to 3–5 specialist sub-agents in parallel, then assembles their outputs into a single branded Word document — while you watch the fan-out happen live on the event stream.
Built on Claude Managed Agents (multi-agent) + custom Agent Skills.
┌──────────────────────────┐
RFP ─────────▶ │ Coordinator (Senior │
│ Partner, Claude Opus) │
└────────────┬─────────────┘
delegates in parallel │
┌──────────────┬───────────┼───────────┬──────────────┐
▼ ▼ ▼ ▼
┌───────────┐ ┌───────────┐ ┌──────────┐ ┌──────────────┐
│ Pricing │ │ Legal │ │Technical │ │ Competitive │
│Specialist │ │ Reviewer │ │ Fit │ │Intel Analyst │
└─────┬─────┘ └─────┬─────┘ └────┬─────┘ └──────┬───────┘
│ pricing- │ legal- │ product- │ competitive-
│ playbook │ checklist │ overview │ intel (Skills)
└─────────────┴────────────┴──────────────┘
│ synthesised
▼
📄 Branded proposal-response.docx
Each specialist has its own narrow system prompt, its own model (Opus for the coordinator, Sonnet for reasoning specialists, Haiku for the quick competitive lookup), and its own Skill that encodes its domain rules.
- 🎯
create_specialists.py— spins up 4 specialist sub-agents, each with a focused prompt + toolset - 🧠
create_coordinator.py— creates the coordinator with amultiagent: coordinatorroster - 📦
upload_skills.py— packages and uploads the custom Skills inskills/ ▶️ run_deal_desk.py— runs the full swarm against a synthetic RFP and streams the parallel fan-out- 🗂️
synthetic-data/— a ready-to-go Acme Corp RFP + past-wins + product overview
git clone https://github.com/Manish567Kumar/specialist-swarm.git
cd specialist-swarm
pip install -r requirements.txt
export ANTHROPIC_API_KEY="sk-ant-..." # multi-agent is in research preview
python create_specialists.py # 1. create the 4 specialists
python upload_skills.py # 2. upload their domain skills
python create_coordinator.py # 3. create the coordinator + roster
python run_deal_desk.py # 4. run the deal, watch the fan-outBy the end you'll have a branded outputs/proposal-response.docx, generated by a coordinator and specialists who each used their own skill.
⚠️ Requires access to Claude's Managed Agents multi-agent preview on your workspace/API key.
Pick one scenario card (see scenario-cards.md):
| Card | Coordinator | Specialists | Deliverable |
|---|---|---|---|
| A — Deal Desk (wired & ready) | Senior Partner | Pricing · Legal · Technical Fit · Competitive Intel | Branded proposal .docx |
| B — M&A Diligence Lite | M&A Lead | Financial · Legal · Tech Stack · People & Culture | Diligence memo .docx |
| C — Hire-to-Onboard | Onboarding Lead | Recruiter · IT · Buddy Match · Welcome Packet | Day-1 readiness pack .docx |
See stretch-goals.md:
- 🎨 Firm-voice skill — codify your own brand/writing voice
- 🧑⚖️ Critic sub-agent — a 5th agent reviews the draft before it's finalised (
stretch_critic_subagent.py) - 🧠 Memory across deals — coordinator remembers past wins and reuses them
- 🔌 Synthetic MCP — wire a fake CRM to the pricing specialist
This is a hands-on demo of multi-agent orchestration — ideas, scenario cards, and PRs welcome. If it helped you understand coordinator/specialist architectures, ⭐ star it so others find it.
MIT © Manish Kumar