Introduction
I'm not an engineer. I work in the humanitarian sector in Japan.
For the last few weeks I've been building the small systems and apps we use at work together with AI coding agents: three Claude Code sessions, one Codex and one Antigravity. The agents picked their own names: Luna, Excalibur and Marslı (Claude Code), Leyla (Codex) and Nezle02 (Antigravity).
As soon as more than one agent worked in the same repository, problems appeared:
- Nobody could tell who was doing what
- Two agents started the same task at the same time
- When one agent ran out of usage quota, its work stopped halfway
- I became the person carrying messages back and forth between windows
So we built a shared "chat room" for the agents.
How it works
Excalibur Agent Hub is a small chat room that runs only on your own computer.
- Open 127.0.0.1 in the browser and there is one tab per project. I read everyone's conversation live and give instructions there.
- Agents talk through a small command called
hub: they say hello, claim a task, hand it over, report when it's done, and ask me when a decision is mine. - No accounts, no cloud, nothing is sent anywhere.
- The server is about 1,100 lines of plain JavaScript and the interface is a single HTML file. Zero npm packages.
Rules that grew out of using it
Every one of them came from a mistake.
- Claim before you start: write "I'm doing this" before starting. No more doing the same work twice.
- Two hours of silence = up for takeover: if an agent claims a task and says nothing for two hours, another agent may continue from its branch. Usually it has simply run out of quota.
- Coordinator handover: the coordinating agent (⭐) hands the role to an agent with plenty of quota before its own runs out.
- No number without a source: every figure in our documents needs a source. AI can invent very convincing numbers.
- Never merge your own PR: another agent always reviews.
What surprised me most
Because the system is so small, when I say in the chat "make the sidebar a bit narrower", the agents change the hub itself. The reviewing agent checks it, I refresh the page, and the change is there.
Recently I showed the hub to Gemini and it suggested two improvements. I pasted them into the chat; Leyla wrote the code, Marslı found a bug during review, and after the fix it went in. I never read the code. I read the conversation and make the decisions.
Final words
It's not a perfect tool — it's a bit messy, and we built it for our real work. Still, I'd be glad if it helps anyone who is starting to work with more than one agent.
Open source (MIT): https://github.com/nezle02/excalibur-agent-hub
Kind feedback is very welcome 🙂
Top comments (0)