Customize final_answer output

[tanliboy]

Is your feature request related to a problem? Please describe.

When using smolagents for multi-step tasks, the final output lacks detail and does not follow a specific format. This happens because final_answer outputs a raw string without customization options.

Describe the solution you'd like

I propose making final_answer a customizable step that allows specifying a format or output instructions (possibly using a different model) while still leveraging the agent’s memory.

Is this not possible with the current options?

I attempted to include output formatting instructions in the system prompt, but it didn’t work well due to the presence of other instructions that the agent also needs to follow.

Describe alternatives you've considered

The other alternative is to make final_answer a dedicated tool for output. However, it would require the model to pass all context and reconstruct the task, which can be overcomplicated.