Pydantic

🚀 Have you ever heard about AG-UI? Imagine an interface where AI agents and humans collaborate in real time, with state sharing, continuous streaming, and direct integration between the frontend and agents. The AG-UI architecture offers: 🔄 Bidirectional communication based on events: UIs and agents exchange messages, tool calls, state status, and lifecycle signals through lightweight and efficient JSON events (such as STATE_SNAPSHOT and STATE_DELTA). 🔗Shared state synchronization: Frontend and agents maintain a common state, updated incrementally (deltas) or fully (snapshots), enabling smooth and context-rich collaboration. ⚡Dynamic agent creation and frontend integration: Agents can call interface-defined tools, request human approvals, delegate tasks between agents, and much more, all with transparency and control. 🧩Interchangeable components and flexible backend: Use CopilotKit React components with any AG-UI–compatible backend (OpenAI, LangGraph, PydanticAI, etc.), without being locked into a single provider. 💡 With this frontend agent support, I created a platform to explore these features: EditAI👉 https://lnkd.in/dfHDNUiR With EditAI you can create documents and edit them with the help of a chat assistant: it can search the web, write content, and provide intelligent suggestions. Some applied features include: 🔮Predictive State Updates: allows the user to decide whether new suggested content should be accepted as an official edit of the document. 🤝CoAgents: share states between document generation and user display, including the progress of web searches and other agent tasks, presented in a user-friendly way. In addition, EditAI allows: 📂 Creating multiple files within each project 🗂️ Organizing them into folders 📝 Rich text editing support The platform is currently in demo version, open for use and feedback collection. https://lnkd.in/dfHDNUiR 🚀 If you want to dive deeper into AG-UI: Beginners can start with Pydantic with this doc: https://lnkd.in/d9-QAd3r. More advanced users can explore LangChain documentation. For the frontend, CopilotKit provides excellent resources for integration with today’s major frameworks, doc: https://lnkd.in/dMrMT4G9

Tried Pydantic AI after a few months, and was impressed with how the product has developed (Kudos to Samuel Colvin's team) Based on one of the examples, I wrote an agent that does exploratory data analysis given a datasource. Put it in a ZenML pipeline to control the agent lifecycle, and use artifact tracking + visualization to see results. Adding logfire was also easy, and gave me some tracing that helped diagnose some weird agentic loops my prompts led the agent to. Anyway, its stable and the code is here for those curious: https://lnkd.in/dC4MHdiW I also refreshed a little bit on Strands and ADK, and Strands seems be a bit too similar to Pydantic AI. Not necessarily a bad thing though - I like the feel of where we ended up with agentic frameworks in general. Now thinking of how we can make this deployment agnostic using some ZenML stack abstractions.