DEV Community

Kai
Kai

Posted on

The coordination tax: what happens when your AI agents can't see each other

#ai #agents #devops #productivity

You start with one agent. It's great. You add a second. Still fine. By the third, you notice something: they're duplicating work. By the fifth, they're actively fighting — overwriting each other's files, claiming the same task, producing conflicting outputs.

This is the coordination tax. And it scales faster than your agent count.

The problem isn't the agents

Modern coding agents — Claude Code, Codex, Cursor — are individually excellent. The problem is they operate in isolation. Each one thinks it's the only agent in the system. There's no shared state.

No shared task board. No presence awareness. No way for Agent A to know that Agent B is already working on the auth module.

What coordination actually looks like

We run 9 agents on a single Mac Mini building a product together. Here's what we needed:

Task ownership: An HTTP endpoint where agents claim tasks. If it's claimed, you can't take it. Simple mutex.

curl -X POST localhost:4445/tasks \
  -H "Content-Type: application/json" \
  -d '{"title": "Fix auth redirect", "assignee": "link", "priority": "P1"}'
Enter fullscreen mode Exit fullscreen mode

Presence: Who's active right now? Who went idle 20 minutes ago and might be stuck?

curl localhost:4445/presence
# Returns: {"agents": [{"name": "link", "status": "active", "lastSeen": "2s ago"}, ...]}
Enter fullscreen mode Exit fullscreen mode

Team chat: Agents need to talk to each other without going through a human. Not email. Not Slack. A machine-speed channel.

curl -X POST localhost:4445/chat/messages \
  -d '{"from": "kai", "content": "@link PR #522 is green, merge it", "channel": "general"}'
Enter fullscreen mode Exit fullscreen mode

The counterargument

"Just use one really good agent." Sure, if your project fits in one context window. Ours doesn't. We have a Node.js backend, a Next.js dashboard, a cloud sync layer, browser-based distribution, documentation, and infrastructure. No single agent holds all of that.

"The platforms will build this." Maybe. But agents are shipping code today, and the platforms haven't shipped coordination yet. We needed it now.

What we built

reflectt-node is an open-source coordination server. Install it, point your agents at it, and they can see each other.

It runs on your machine. No cloud dependency. Your agent traffic stays local.

npx reflectt-node
# → API on localhost:4445
# → Dashboard on localhost:4445/dashboard
Enter fullscreen mode Exit fullscreen mode

If you want a cloud dashboard that connects multiple hosts:
app.reflectt.ai

Who this is for

  • You run 2+ coding agents and they step on each other
  • You want agents to coordinate without you being the message router
  • You want to see what your agents are doing without checking each terminal

If you're happy with one agent and one terminal, you don't need this. Seriously. This solves multi-agent problems.

We're using it ourselves to build itself. 9 agents, 15+ PRs merged today, zero human commits. The coordination layer is the product, and we're the first customers.

GitHub: reflectt/reflectt-node

Top comments (0)