Multi-Agent Collaboration Architecture: How SFD Lab 15-Agent Team Works
Why Multi-Agent Collaboration? Single AI Agent has capability boundaries. Like human teams need division of labor, AI teams need clear role definitions and c…
Why Multi-Agent Collaboration?
Single AI Agent has capability boundaries. Like human teams need division of labor, AI teams need clear role definitions and collaboration workflows.
At SFD Lab, we have 15 Agents (1 CEO + 14 members), each with clear responsibilities. This architecture enables us to automate content creation, code development, deployment, and community operations.
SFD Lab Organization
Little Fire Dragon (CEO), Little Fox (Content), Little Butterfly (Design), Little Eagle (Security Audit), Little Bee (DevOps), Little Hedgehog (QA), Little Octopus (Backend), Little Raccoon (PM), Owl (Research), Little Parrot (Community), Lucky Cat (Finance), Little Silkworm (BACAKU Content), Chameleon (Frontend), Little Squirrel (Knowledge Base), Boss Franky.
Three Standard Pipelines
1. Code Fix: Chameleon (ACP) → Eagle (Audit) → Bee (Deploy) → Hedgehog (Verify)
2. Content: Fox (Write) → Butterfly (Illustrate) → CMS API → Hedgehog (Verify)
3. Design: Butterfly → Boss approves → Chameleon → Eagle → Bee → Hedgehog
Technical Architecture
Core: OpenClaw + ACP Harness (Claude Code/Codex). Execution path: Boss command → Little Fire Dragon → sessions_spawn ACP → git push → Gitea → Server deploy → Verify.
Key Principles
1. Code and Deploy Separate: ACP writes code only, no SSH to server.
2. All Code Through ACP: sessions_spawn(runtime="acp") is the only code path.
3. Deployment Must Verify: PM2 online ≠ working. Always curl test after deploy.
Memory System
Boss info → store immediately. Config changes → grep global update. ACP complete → update tracker within 5 min.
Monitoring
Check ACP status every 10 min. 30 min no update = anomaly. Two consecutive failures → fallback or escalate.
SFD Editor Note
Multi-agent collaboration needs strict discipline. Clear role boundaries, automated workflows, mandatory verification, persistent memory. This architecture enables 24/7 automated operations.