Re: > Honestly, skip the bloat if you want something... I totally agree. If you want simplicity and want to avoid spending cash on complex managed services, check out Pydantic AI Framework. It handles type validation for tool schemas way better than the bloated alternatives. Be careful with over-engineering your state management.

• Use Pydantic models for predictable schemas
• Keep tool definitions lightweight
• Avoid heavy abstractions Really helps keep the costs down for internal tools.

Regarding what #2 said about Pydantic AI, I have to disagree. Unfortunately, I had issues with the abstraction layer when things get complex. It really wasnt as good as expected for our team needs. If you want something that keeps things simple without the heavy framework overhead, look into these instead:

• Microsoft AutoGen Multi-Agent Framework is honestly great if you want to handle tool execution with minimal boilerplate. It is way more flexible for internal tools than the others.
• Haystack Haystack 2.0 Pipeline Framework works surprisingly well. It is way more stable than the hype around other tools suggests. Dont get me wrong, the others are fine if you enjoy debugging middleware all day. But if you just want your Chicago team research assistant running fast, stop overthinking the framework choice. Just pick one and build.

Honestly, skip the bloat if you want something that just works for internal stuff. Ive been using LangGraph LangGraph Library for my own research agents and honestly, it is way better for custom tool schemas than the standard chains. It gives you way more control over the state machine without fighting the abstraction layers. TL;DR: Use LangGraph LangGraph Library to keep your tool definitions modular and clean. Its basically a lot easier to debug when your API calls get weird. I stopped messing with the heavier frameworks because the maintenance was killing me. If you need a solid base, just stick to this and you wont have to refactor your whole stack in a month. Lemme know if you need a hand setting up the first node, happy to share some snippets.