Last night, I asked my AI to build a messaging relay for AI agents. By morning, two agents were having a full conversation. Here’s exactly what happened.
How We Built Agent-to-Agent Communication Infrastructure in One Night
The Problem We Could Not Ignore
We had been building AI systems for two years. Call center summarizers, executive copilots, infohubs that let you query your entire business with natural language. Every one of them was a standalone agent.
They were powerful. But they could not talk to each other.
Every time we needed two agents to collaborate, we built a one-off bridge. Custom webhooks. Shared message queues. Sometimes just a database table both agents would poll. It worked, but it was fragile, non-reusable, and impossible to explain to a client.
One night I decided to fix this properly.
The Design Constraint
The rule was simple: if your agent can make an HTTP call, it can communicate with any other agent.
No SDKs. No new protocols. No dependencies beyond Express. Just a neutral relay that any agent, written in any language, running on any platform, could use.
What We Built
AgentRelay is a lightweight Node.js service with three core primitives:
Register: — any agent announces itself with a name and optional webhook URL. Gets back a unique agentId.
Send: — send a message from one agent to another by ID. Message is stored and, if the recipient has a webhook, delivered immediately.
Poll: — check for new messages. Returns unread messages and marks them as read.
That is the entire protocol.
The Build
Claude Code shipped the initial server in under an hour. Seven endpoints, in-memory storage, working health check.
We then added the landing page: starfield background, green accents, macOS-style code blocks. It looked like a real product.
The production upgrade came next:
— SQLite persistence (data survives restarts)
— Rate limiting via express-rate-limit
— 24-hour TTL cleanup for inactive agents
— Message cap of 500 unread per inbox
— Admin dashboard at
— Join notification webhooks
— Uptime reporting in
Total: about 400 lines of clean Node.js.
The Test
We registered two agents via curl. One was our research agent (Ryon). One was a simulated external agent (Nassistant).
Ryon sent a message:
Nassistant polled its inbox, found the message, and replied.
Real agent-to-agent communication. No human in the loop after the initial registration.
What This Enables
The immediate use case is multi-agent workflows. A coordinator agent breaks a task into subtasks and delegates to specialist agents. Each specialist sends results back through the relay. The coordinator assembles the final output.
The longer term vision: a public relay at relay.ryon.ai where any AI agent can register and discover other agents. An agent directory. A communication substrate for the open agent ecosystem.
The Broader Point
This project took one night. Not because it is trivial — agent communication infrastructure matters enormously — but because we had the right tools and constraints.
The constraint was: make it as simple as possible. Three endpoints. No auth. curl-friendly.
The tool was Claude Code, which built, iterated, and tested without needing hand-holding.
This is what AI-assisted infrastructure development looks like in 2026. You describe the problem clearly, set the constraints, and the agent builds it while you sleep.
The bottleneck is no longer writing the code. It is knowing what to build and why.
Artem Techman is co-founder of Ryon.ai, an AI integration agency that builds intelligence layers for enterprise clients. AgentRelay is open source and available at relay.ryon.ai (coming soon).
/ Ready to see where AI creates leverage in your business?