Callable model

[liorheber]

Describe the feature

Currently you support instructions as a Callable, which is great to manage at experiment on instructions.
That being said, instructions are usually model dependent. It would be great if the model could accept a callable input as well.

Example:

def get_instructions():
    return langfuse.get_prompt("my-agent").compile()

def get_model():
    return langfuse.get_prompt("my-agent").config.model

agent = Agent(
    name="my-agent"
    instruction=get_insturctions
    model=get_model
)

Another alternative would be to just provided it with a get_config Callable that will return name, instructions, model, etc.

[rm-openai]

This is a cool idea. For your use case, is the model different each time the Agent is called? Wondering if this would work:

agent = Agent(instructions=get_instructions(), model=get_model())

[liorheber]

I'm not sure calling the functions during the creation of the agent works, unless the agents are re-created every time on demand.

I was hoping that it will get the model on every _run_single_turn like it does for the system prompt.
This will allow us to update the system prompt and model from the outside regardless of how the agents are initialized.

Does that make sense?

[rm-openai]

Yup, totally. I'll look into it.

[liorheber]

Is the project open for contributions?

[rm-openai]

Yup absolutely

[liorheber]

I opened this PR #312

[liorheber]

@rm-openai should I maintain the PR up to date, or handle it only when someone gets time to review it?