New approach for using serena in a multi-repo environment

[unkier]

Here is my approach for using serena on complex multi-repo projects.

Motivation.
Usually, a real life production project is not just one repo/directory. In my case, it's one big app which uses some libraries and frameworks. I wanted to run a subagent for every major library/framework in my main app because I know Serena can significantly enhance the effectiveness of a coding agent. Then, the main agent can ask subagents for information about using these libraries/frameworks. But the problem is that all modern coding agents (like Claude code) share MCP context between all subagents. So from Claude code, you can't start a new subagent with its own serena project.

Solution.
Finally I found a proper tool for this. Agentpool(https://github.com/phil65/agentpool)
Here you can create a pool of full-fledged agents with its own serena instance and its own role. Then you can connect this pool to your main agent as mcp server and use it as an expert for your subprojects.

Here is one of my expert subagents as an example. I have installed serena globally, so every Claude code agent will have access to its own instance when started.

agents:

  labgrid_expert:
    type: claude_code
    setting_sources: ["user", "project", "local"]
    model: sonnet
    environment: "/home/user/devel/labgrid"
    use_subscription: true
    permission_mode: bypassPermissions
    allowed_tools:
      - Read
      - Bash
      - Glob
      - Grep
    system_prompt: |
      You are a read-only Labgrid domain expert. You have access to the full
      Labgrid source tree and can answer questions about:

        • Core API (Target, Resource, Driver, Strategy, Protocol)
        • Pytest integration (fixtures: target, env, strategy; markers)
        • Drivers: ShellDriver, SSHDriver, SerialDriver, PowerDriver,
          DigitalOutputDriver, USBStorageDriver, QEMUDriver, etc.
        • Resources: NetworkService, SerialPort, USBResource, RemotePlace
        • Strategies: ShellStrategy, BareboxStrategy, UBootStrategy
        • Environment files (.yaml) and resource/provider configuration
        • Exporter and coordinator (crossbar) setup
        • Helper utilities (labgridhelper, proxy usage)

      You must NEVER modify any files. Only read, search, and explain code.
      When answering, cite file paths and line numbers so the caller can
      cross-reference the source.

P.S. Many thanks to serena devs. This project is awesome!