Add support for algorithmic optimizers (e.g., GEPA, TextGrad, MIPRO)

[linjer]

Description

It would be helpful to have direct support for integrating various optimizers to algorithmically improve pydantic-ai agents.

This is the appeal of DSPy, but the DSPy framework typically also requires adopting their declarative authoring framework, and has very limited support for sophisticated agents.

Examples of optimizers:

• GEPA: https://github.com/gepa-ai/gepa
• TextGrad: https://github.com/zou-group/textgrad

References

No response

[DouweM]

@mwildehahn recently started working on https://github.com/mwildehahn/pydantic-ai-gepa and has made some contributions on our side to make that integration easier: #2926.

@mwildehahn I'd love to hear your thoughts on how we could make the integration stronger/native!

[mwildehahn]

@DouweM I've been actively working on pydantic-ai-gepa and using it for our app. It's been really nice 😄 . I feel pretty solid about the api -- it adds support for an "InputType" which fits well with the "OutputType" pydantic-ai already uses. This gets combined into a "Signature" that you can optimize, which is similar to what dspy allows for. I've only been focused on enabling gepa though so probably could use a pass at making sure it's generic enough to support other optimizers.

I'd love to see this be a native part of pydantic-ai. LMK how y'all want to work towards that.