Adding/removing a tool should also update the system prompt accordingly

[andrewrreed]

Currently, if the user updates agent.tools, say to add or delete a given tool, that update is not reflected in the agent.system_prompt.

I recall there used to be dedicated functionality for managing a toolbox, but with the port to smolagents I think this got lost

[AngeloKiriakoulis]

The agent.system_prompt does update to reflect changes in agent.tools, but only after the agent performs a step. During steps, tools are reconfigured and re-parsed, which triggers the system prompt update. Inside step:

try:
    model_message = self.model(
        self.input_messages,
        tools_to_call_from=list(self.tools.values()),
        stop_sequences=["Observation:"],
    )

For example:

agent = ToolCallingAgent(tools=[get_weather], model=model)
print(agent.system_prompt)  # Lists 'get_weather', 'final_answer'

agent.tools["get_news"] = get_news
del agent.tools["get_weather"]
print(agent.system_prompt)  # No change yet

agent.run("what is the weather in Paris? Also inform me about the news there.")
print(agent.system_prompt)  # Now lists 'final_answer', 'get_news'

[aymeric-roucher]

You're right @AngeloKiriakoulis ! The logic is handled in initialize_system_prompt() which runs at the beginning of each MultiStepAgent.run() operation.

This is because system prompt is more agent logic than toolbox-logic, so changing system prompt on the toolbox level would probably have created some broken expectations/lower readability for the user. (it was already the case in transformers.agents @andrewrreed )

[aymeric-roucher]

I think the current state has good functionality for low complexity (especially given the distinction between self.system_prompt and self.system_prompt_template, which makes it clearer that the system_prompt will be derived from system_prompt_template), but tell me if that is otherwise!

[andrewrreed]

This makes sense to me, thanks for the clarification both!