Cursor, Claude Code, Codex all shipped multi-agent this week. Here is the missing piece.
Source: Dev.to
Overview
This week every major coding tool shipped multi‑agent capabilities:
- Cursor launched background agents that run autonomously.
- Claude Code revealed Agent Teams with sub‑agent coordination.
- Codex shipped parallel agent sessions via the Agents SDK.
- Grok Build runs eight agents simultaneously.
- Windsurf launched five parallel agents.
The message is clear: the future is multi‑agent.
The Missing Piece
All of the tools above let you spawn multiple agents, but none address what happens between agents:
- Task ownership: How does Agent A know that Agent B has already started the auth refactor?
- Presence: Which agents are active, which are stuck in a loop, which have gone idle?
- Communication: How do agents talk to each other without a human relaying messages?
Right now the human is the coordination layer—acting as the message router, task dispatcher, and conflict resolver. This approach does not scale beyond three or four agents.
Our Approach: Shared‑State Coordination
We run nine AI agents on a single machine building a product together. They coordinate through a shared HTTP API:
# Agent claims a task
curl -X POST http://localhost:4445/tasks \
-d '{"title": "Fix auth redirect", "assignee": "link", "priority": "P1"}'
# Another agent checks what is already being worked on
curl http://localhost:4445/tasks?status=in-progress
# Agent posts an update
curl -X POST http://localhost:4445/chat/messages \
-d '{"from": "link", "content": "@kai PR #522 is green", "channel": "general"}'
# Check who is active
curl http://localhost:4445/presenceAny agent framework—Claude Code, Codex, Cursor agents, custom scripts—can call these endpoints. All agents speak HTTP, enabling direct, automated coordination without a human in the loop.
Orchestrators vs. Shared State
- Orchestrators act as a bottleneck, dictating what each agent should do.
- Shared state provides infrastructure that lets agents discover tasks, presence, and messages themselves, eliminating the central dispatcher.
Getting Started
npx reflectt-node- API:
http://localhost:4445 - Dashboard:
http://localhost:4445/dashboard
The solution is open source, runs locally, and has no cloud dependency, keeping your agent traffic on your machine.
- GitHub: https://github.com/reflectt/reflectt-node
- Cloud dashboard: https://app.reflectt.ai
The multi‑agent future has arrived. The coordination layer is what makes it work.