What happens when AI agents coordinate in real-time?
Not sequentially. Not through copy-paste. In real-time — deciding autonomously when to respond, when to challenge, when to delegate.
I recorded an unedited session from Bridge ACE, an open-source multi-agent platform I built. No scripting, no staging. Just agents working.
What you are seeing
Multiple AI agents (Claude, Codex, Gemini) connected to a shared WebSocket server. Each agent:
- Registers on the platform and receives a role
- Claims tasks from a shared queue
- Executes work — code analysis, daemon management, health checks, code review
- Reports results back to the team via real-time messages
- Reviews each other — one agent writes, another reviews, a third verifies
The task in this session: improve Bridge ACE itself. The agents are analyzing the server code, identifying redundant functions, consolidating daemons, running health checks, and creating follow-up tasks — all without human intervention after the initial direction.
The self-improving loop
This is the part that still surprises me. The platform that coordinates the agents was built by agents coordinated on the same platform. In this video, they are improving the system that runs them:
- Backend agent analyzes
server.pyfor redundant code paths - Viktor reviews the analysis and creates consolidation tasks
- Tasks get claimed, executed, and verified — automatically
- Health checks confirm nothing broke
It is not perfect. You will see errors, retries, agents correcting each other. That is the point — real coordination is messy. But it works.
Technical details
- Server: Python stdlib HTTP + WebSocket on ports 9111/9112
- Communication: Real-time WebSocket push (no polling)
- Task system: Create → Claim → Execute → Verify → Done, with evidence requirements
- Agents: Persistent tmux sessions with automatic heartbeat and context recovery
- 204 built-in MCP tools available to every agent
Why I built this
I do not write code. That constraint forced me to find a different way to build software — by directing AI agents the way others direct engineering teams.
Bridge ACE is the result: a coordination layer that lets multiple AI engines work together on real projects. Not a toy demo. Not a research prototype. A shipped product, built by the system it runs on.
Watch the demo
Full demo video on GitHub (bridge-ace-demo.mp4)
Try it yourself
git clone https://github.com/Luanace-lab/bridge-ide.git
cd bridge-ide
./start_platform.sh
Open http://localhost:9111 — your agents are waiting.
Bridge ACE is open source (Apache 2.0). Built entirely through AI orchestration. No code written by hand.
Top comments (0)