pydantic/pydantic-ai ? reverse-engineered prompt

Hey, I'd like to build a Python framework for creating AI agents. The main idea is to bring the developer experience of FastAPI to the world of generative AI. It should be heavily based on Pydantic for data validation and type safety, so my IDE can help me out and I catch errors early.

It needs to be super flexible. Let me plug in any model I want—OpenAI, Claude, Gemini, even local ones via Ollama. The core concept should be building agents from composable "capabilities." For example, I should be able to easily add web search or a "thinking" module to an agent.

For more complex tasks, let's support defining agent workflows as a graph. It also needs great observability out of the box, maybe integrating with OpenTelemetry, so I can trace and debug everything. Please also add a system for evaluating how well the agents are performing and a "human-in-the-loop" feature so I can approve certain tool calls before they run.