Feature request: fallback agents and sub agents

[rdehuyss]

I'd like the possibility to configure multiple agents, both for fallback purposes and sub agents for smaller tasks like summarization to reduce costs.

[arefbehboudi]

I would be happy to work on this.

Before starting, I would like to propose an initial approach and get feedback to make sure it aligns with the project direction.

My idea is to first introduce proper multi-agent support, without adding sub-agents. The goal is to move from a single DefaultAgent to a more extensible architecture where multiple agents can coexist with isolated configuration and runtime state.

Proposed approach

1. Agent abstraction
   Instead of relying on a single DefaultAgent, introduce:

• AgentDefinition (id, name, system prompt, tools, model/provider, memory scope)
• AgentRuntime (wrapper around ChatClient + tools + memory)
• AgentRegistry (to manage available agents)

This keeps agent configuration declarative and makes it easy to add/remove agents.

2. Agent-aware routing
   Add a lightweight routing layer to determine which agent handles a request:

• explicit selection (UI switch)

Initially, I would keep this simple to avoid overcomplicating the first iteration.

3. Memory isolation
   Ensure chat memory is scoped by:

• agentId
• conversation
• channel

This prevents context leakage between agents and makes behavior predictable.

4. Per-agent configuration
   Each agent should be independently configurable:

• system prompt (e.g. separate AGENT.md per agent)
• model/provider (aligned with the pluggable provider direction)
• enabled tools

This also allows different agents to use different models or capabilities.

5. Tool scoping
   Tools should be injected per agent instead of globally, so each agent only has access to what it needs.

6. Job / task awareness
   Since JobRunr is used for background execution, tasks should include agentId in their metadata so that execution context is preserved when jobs run asynchronously.

[rdehuyss]

Hi @arefbehboudi ! Thanks for your help on this. Thanks to @tzolov, I just found out that Spring also has something for multiagent already. There is a pointer to an example here: https://github.com/tzolov/voxxeddays2026-demo/tree/main/16-agents-subagent.

What do you think of this approach?

[arefbehboudi]

@rdehuyss Thanks, this is a great pointer.

This looks like the right direction and keeps JavaClaw aligned with Spring AI primitives. The example models a delegation-based setup, where a main agent coordinates specialized sub-agents via TaskTool, with isolated prompt/context/tool scopes and per-agent model routing.

One distinction is that fallback (failover/retry across agents or models) is a separate concern from this delegation model, so it would likely need to be layered on top.

If that sounds aligned, I can explore this approach and propose a JavaClaw integration.

[arefbehboudi]

Hi @rdehuyss,
I have created this merge (#48) as the foundation for multi-agent support. It is not fully complete yet, but it sets up the core structure and direction.

What do you think of this approach?

[rdehuyss]

Thanks, I will review tomorrow!

[arefbehboudi]

Thank you a lot! It is not fully complete yet and still has some bugs. I will try to fix them before your review.