Multi-Agent Orchestrator Mode

[jakeschurch]

I've seen other editors support workflows where AI agents collaborate to solve tasks. I haven’t found any mention of an “orchestrator” or multi-agent mode in current discussions.

I absolutely love CodeCompanion (CC) - ty to OliM and the community for building such a delightful plugin! I'm exploring building a proof-of-concept (either as a plugin or an CC extension) to add multi-agent orchestration. It feels doable given what I think I know, but I’d really appreciate help validating my assumptions before diving in deeper.

Goals

• Build a lightweight orchestrator enabling multiple agents to work in concert—e.g., "lint -> test -> refactor."

• Reuse in CC’s framework with minimal duplication.

• Validate implementation assumptions before investing time.

Current Assumptions / Feedback Request

1. Agent Creation

Each agent would spawn its own dedicated buffer for handling its messages, chat context and state.

2. Shared Chat Context for Orchestration

Maintain a shared “chat message bus”, where agents listen and post.

Example syntax in the bus could be something like:

$lintAgent: fixed whitespace
$testAgent: generated 10 tests

3. Buffer Roles & Message Routing

The orchestrator would need to route messages between agents.

I’m uncertain about: What role should be assigned when sending one agent’s output as another agent’s input?

4. Agent UX

As I've delved into the creating tools doc, it does not seem like I would be able to currently create a CC extension, would you agree?

Do you have any thoughts about how I should proceed with such a POC?

Please let me know if I am missing anything - like being able to potentially turn user messages into agent workflows - and having custom subscribers / watchers to alert / update other buffers, thanks!

[olimorris]

This is a very, very cool idea. Something I would love to include in the plugin at some point. I'm been planning on adding a background strategy to allow external plugins to leverage CodeCompanion in the background and mooted the idea of adding a call_llm tool to allow LLMs to leverage the strategy and call other LLMs in the background. A basic way of having an agent orchestrator in the chat buffer.

The next couple of months I'm planning some core functionality around modes (like GitHub Copilot Chat) and the background strategy. Once those are over, the plugin will be in a pretty awesome state. But, that shouldn't hold you back. I think there's a case that what you've proposed is better being in an extension (external plugin). I expect there would need to be some modifications to core functionality in CodeCompanion but I'd be happy to work with you on that.