Question about subagent resolution with same input type

[pweemeeuw]

I’m experimenting with subagents and noticed something odd compared to the example (https://github.com/embabel/embabel-agent-examples/blob/main/examples-java/src/main/java/com/embabel/example/handoff/Subagents.java).

In my own setup, where both subagents accept the same input type (UserInput), only the last subagent seems to be invoked and produces an answer, while I expected that the described goals would be used to route to the correct agent:

operationContext.promptRunner()
    .withLlm(LlmOptions.withAutoLlm())
    .withSubagents(
        Subagent(subAgentBasicRequests, UserInput::class.java),
        Subagent(subAgentFilteringRequests, UserInput::class.java)
    )
    .createObject("The user searches for models and asks: ${operationContext.input}", ModelResponse::class.java)

Am I missing a configuration to let both subagents contribute?
Any clarification on how subagent resolution works in this case would be much appreciated.

[johnsonr]

withSubagents works via exposing agents as tools--what other frameworks call "handoff". Thus the LLM will choose which one to invoke and you're out of the world of GOAP planning.

I think we need to support handoff, but it is inherently deterministic.

Subagent support is presently experimental and not well documented, but hardening it is scheduled for 1.0. #870

[igordayen]

@pweemeeuw - could you please update? Do you still experience an issue in release 1.0.0? thank you

[pweemeeuw]

Hi Igor, I retested with Embabel 1.0.0. The LLM correctly selected the agent, so the issue is resolved. Thanks!

[igordayen]

@pweemeeuw Happy to hear that. Thanks for the update!