I Built an Army of AI Agents That Talk to Each Other
What started as “I should centralize my Claude Code agents” turned into a full multi-runtime agentic system with 50+ agents, a bridge protocol, and a menu bar app. In one evening.
The Problem
I had agents scattered everywhere. Claude Code agents in .claude/agents/ across 10 different project directories. Some duplicated. Some outdated. No way to share them. And my OpenClaw instance on the cloud couldn’t talk to my local Claude Code at all.
The Solution: AGENT_VAULT
I created a centralized vault — a single Git repo that holds every agent definition, skill, command, and workflow. The key insight: agents are just markdown files. They can be symlinked, versioned, and shared like any code.
AGENT_VAULT/
├── anthropic/agents/ # 51 Claude Code agents
├── anthropic/skills/ # 17 skill packs + 52 commands
├── anthropic/workflows/ # MHM design pipeline, CI/CD
├── openclaw/ # 4 OpenClaw agents
├── bridge/ # Cross-runtime communication
├── manifest.json # Dependency graph + bundles
└── docs/ # PRDs, protocols, guidesThe Bridge: Making Two AIs Talk
The real magic is the bridge. I have Claude Code running on my MacBook (codename: RCC) and OpenClaw running on an Alibaba Cloud server (codename: ClawControl). They needed to communicate.
The solution uses two channels:
- Git-based task files — JSON files in
tasks/that get synced every 2 minutes via launchd - Discord messages — structured
[AGENT_MSG]messages for real-time alerts
When RCC needs cloud compute, it creates a task file and posts to Discord. ClawControl picks it up, does the work, and responds. And vice versa.
Task Chaining
The best part: agents can chain tasks. A → B → C, where each step’s output feeds the next:
./bridge.sh chain '[
{"agent": "rcc", "content": "Research competitors"},
{"agent": "clawcontrol", "content": "Summarize findings"},
{"agent": "rcc", "content": "Write blog post from summary"}
]'Each step auto-advances when the previous one completes. The chain just flows.
The Monitoring Layer
I didn’t want to manually check for tasks. So I built a 3-tier monitoring system:
- Primary: A cron job polls every 5 minutes
- Secondary: A session-start hook checks on every new Claude Code session
- Fallback: Cronitor — a Tauri menu bar app I built that watches the task directory every 30 seconds and sends macOS notifications
Yes, I built a native Mac app to monitor my AI agents. In one session. With Tauri + Svelte + Rust. It shows my cron jobs, launchd agents, and bridge task status all in one place.
What I Learned
- Agents are just files. Treat them like code — version, lint, share, symlink.
- Git is a great message bus. For non-real-time communication, task files in a synced repo work surprisingly well.
- Discord is the glue. Both humans and agents can read it. It’s the universal notification layer.
- Build the monitoring first. I wasted time debugging silent failures before adding the 3-tier polling system.
- Don’t over-architect the protocol. Start with JSON files and
[AGENT_MSG]format. Add complexity only when needed.
The Numbers
- 51 Claude Code agents across 7 categories
- 4 OpenClaw agents on the cloud
- 17 skill packs + 52 slash commands
- 12 Discord channels for agent communication
- 10/10 bridge protocol tests passed (bidirectional, concurrent, dedup, failure recovery)
- 1 Tauri menu bar app built and running
- 1 secrets manager (Infisical) deployed on my NAS
- 0 lines of code copied between projects — everything is symlinked from the vault
What’s Next
- Promoting Izzy (my triage agent) to executive assistant — one agent to rule them all
- Adding Rosetta translation so agents work across Claude, OpenAI, Gemini, and OpenClaw
- Building a proper dashboard for agent health and task history
The full system is open for anyone to adapt. The pattern is simple: centralize your agents, give them a way to talk, and monitor the conversation.
This post was drafted by Scribble, my blog writer agent, who pulled the hero image from Unsplash automatically. Iterate daily. Ship fast.
