Theia AI - Multiple AI Agent collaborating with each other

[fayakun-it-consulting]

Hi to the community!

I became aware of the possibility of defining custom AI Agent within Theia AI which is a pretty neat feature.
However I didn't see any tutorial showing us how to implement a multi-agent collaboration with functionalities such as:

1. Tell a specific agent to go on a web site to read all the documentation for a specific project.
2. Once the ingestion of the documentation finished put that information into a memory database (maybe RAG)?
3. Render that information available for the other agents for collaborative tasks

Is this possible right now ?

[planger]

Hi @fayakun-it-consulting,

Yes, multi-agent collaboration is possible in Theia AI since version 1.63 🎉

It now supports agent delegation via a tool call:
delegateToAgent("AgentName", "Prompt")

That way, one agent can delegate a task to another and get its response as part of its own context. Here's what you'd add to agent A's system message to enable that:

## Delegation
You can delegate tasks to other agents with this tool:
- **~{delegateToAgent}**

Agents available for delegation are:
- `Architect`: to analyse the current code base and design new features
- `Coder`: to implement code

Agent B can even delegate again if needed. The delegated agents don’t get access to the full chat session, only the prompt passed to them; so the delegating agent (agent A) can fully control the context for agent B.

By default, the delegation prompt and response will be shown in a collapsed panel in the chat view, but you can customize that by overriding the corresponding response part renderer.

Changes made by agents are bubbled up to the main change set, so multiple agents can work collaboratively on workspace changes.

Memory or ingestion isn’t built-in yet, but could be integrated by giving agents access to appropriate tools.

Hope that helps!

[planger]

Absolutely! There is already a variable called #chatAgents that you can use in your prompts to resolve all registered agents. So you could put that into the system message when instructing the LLM which agents are available.

One caveat: It certainly depends on the use case and environment, but our experience is that it works best if you also give some hints on when to delegate which type of task to which agent. Otherwise the results can quickly get very unpredictable with many available agents and no further instructions.

[planger]

PS: See https://github.com/eclipse-theia/theia/blob/master/packages/ai-chat/src/common/chat-agents-variable-contribution.ts