October 20, 2026

In June I wrote about Multica, an open-source platform to manage coding agents as members of a team, and said I wanted to try it. I have been using it since, and it is now part of how I maintain some of my projects.

The setup

Multica runs on the home server I mentioned in the post about UDP hole punching. The agents behind it are the same ones I was already using separately: Claude Code, Codex and OpenCode with Qwen running locally. The difference is that now they are coordinated from one place instead of me switching between them by hand.

Squads and roles

The agents are organized in squads, and each squad works on a specific project. I set them up as small multidisciplinary teams, where every agent has a defined role:

• Scout and Team Leader: looks for issues in the project and coordinates the work.
• Analyzer: takes an issue and writes its description.
• Builder: checks that the description holds up, then implements the fix.

The Scout is also the one that moves the work forward. When a step is done, it changes the state of the issue and reassigns it to the next agent, so the issue goes from discovery to analysis to implementation without me having to pass it along.

Splitting analysis and implementation between two agents means that the Builder does not start from a raw finding but from a written description, and the first thing it does is verify it.

Where I come in

The flow ends with a pull request on GitHub, and nothing gets merged until I approve it. In the meantime I can step in on any issue directly in Multica, to correct it, add context or stop it.

What it is good at

For now it works very well on small projects that I want to improve a little at a time. This site is one of them. Every day the flow starts on its own and the squad fixes small things, from layout problems to routing errors. None of these would justify setting aside an afternoon, but together they add up, and the work arrives as a list of pull requests to review.

What I wanted to verify in June was whether a single coordination layer over different agents is worth it. For this kind of continuous, incremental maintenance, it is.