💡 MAEA is a governance framework and reference architecture — protocol specs, governance models, and Agent design methodology. Open for discussion and iteration. Join the Discussions.
When you have more than 5 AI Agents, MAEA provides communication protocols, governance frameworks, and quality validation — turning your Agents from "a pile of prompt templates" into "a governable AI team."
agency-agents earned 125k stars with 232 Markdown files. But open the Issues tab — the top-voted items aren't "add more Agents":
- "Which of these agents actually work?" (#11, 17 👍)
- "I had to build my own orchestration engine" (#460)
- "I can't choose among 232 agents" (#634)
The demand is validated: it's not more Agents people need — it's governance.
→ Deep dive: Why agency-agents went viral but wasn't enough
┌─────────────────────────────────────────────────────────┐
│ Human (Feishu DM) │
└──────────────────────┬──────────────────────────────────┘
│
┌────────▼────────┐
│ DM Hub │
│ sg-architect │ Routing + Governance
│ :9900 │
└───┬──────┬──────┘
│ │
┌─────────────┤ ├─────────────┐
│ │ │
▼ ▼ ▼
```┌────────┐ ┌──────────┐ ┌──────────────┐
│Research│ │ Delivery │ │ Management │
│:9920 │ │:9912:9910│ │ WLM │
│ │ │ dev ops │ │ Scheduler │
└────────┘ └──────────┘ └──────────────┘
│ │ │
└─────────────┼───────────────────┘
│
┌────────▼────────┐
│ Tools (MCP) │
│ search · files │
└─────────────────┘
Three buses, five layers, 8 Agents running on Feishu today.
Human ──Feishu──► DM Hub (sg-architect) ──A2A──► Agent A
│ Agent B
│ Agent C
└────MCP────► Tools (search/files/...)
Three Communication Buses:
| Bus | Protocol | Purpose |
|---|---|---|
| 🤝 Agent ↔ Agent | A2A | Delegation, review, conflict arbitration |
| 🔌 Agent → Tools | MCP | Data retrieval, patent analysis, search |
| 👤 Human → Agent | Feishu | Single-interface routing via DM Hub |
Five-Layer Governance Model:
Strategy — sg-architect Architecture governance + routing hub
Research — cm-deepsight Deep research + red-team argumentation
Delivery — do-developer Code implementation / do-ops operations
Management — WLM Workload management (IBM z/OS WLM style)
Gateway — gw-default Feishu / A2A Gateway
→ Further reading: Framework Overview · Protocol Specs
| Entry Point | Time | What You Can Do |
|---|---|---|
| Claude Code | 30s | A governance Agent inside Claude Code — picks Agents, detects conflicts |
| Hermes | 30min | Real A2A network, inter-Agent comms + Feishu integration (reference impl) |
| Read only | 0min | Methodology + protocol specs |
git clone https://github.com/deeparchi-ai/MAEA-Framework.git
cd MAEA-Framework/demo
python3 demo.pyOutput:
Phase 1: Registry ✅ 3 agents registered
Phase 2: A2A Ping ✅ all 3 reachable
Phase 3: Delegate ✅ task routed correctly
Phase 4: Conflict ✅ overlap detected
4 phases validating A2A protocol + registry + routing + conflict detection. Zero external dependencies.
| You | Start Here |
|---|---|
| 🤖 Solo dev, 3–5 Agents | Claude Code skill → experience governance first |
| 👥 Team, 5+ Agents | Hermes reference impl → A2A protocol + registry |
| 🏗️ Agent platform/framework builder | specs/ → A2A protocol + registry spec |
| 📖 Here for the methodology | docs/ → governance + Agent design + quality validation |
MAEA is a framework, not a product. Sub-projects:
| Project | Description |
|---|---|
| macs | Multi-Agent Coordination Runtime |
| wlm | Goal-driven resource scheduler (IBM WLM style) |
MAEA is open source. If you're also drowning in ungoverned Agents — let's build the governance layer together.
125,000 people bookmarked an AI dream team. The next question isn't "what other Agents can we add?" — it's "who's in charge?"