Summary
Pydantic AI 2.0 represents a significant architectural rethink of one of Python’s most popular AI agent frameworks, and Cole Medin — who has covered Pydantic AI since January 2025 — provides a thorough walkthrough of the changes. The centerpiece of the 2.0 release is the “capability,” a new composable primitive that bundles an agent’s instructions, tools, lifecycle hooks, guardrails, and model settings into a single reusable unit. Medin describes this as moving from a loose bag of parts to structured Lego blocks that can be snapped together and shared across agents and projects.
The video contrasts the 1.0 approach — where system prompts, tools, and model settings were defined as a flat collection inside the agent constructor — with 2.0’s capability composition pattern. A live GitHub repository demonstrates both versions side by side, and Medin positions capabilities as a layer above MCP servers: while MCP packages tools for portability, capabilities go further by encapsulating behavioral intent, guardrails, and human-interaction hooks alongside the tooling.
Medin also addresses the competitive context directly, noting that the rise of coding agent SDKs from Anthropic (Claude Agent SDK) and OpenAI (Codex SDK) had been quietly narrowing the use cases where a dedicated framework like Pydantic AI made sense. His argument is that the 2.0 capability model reasserts the framework’s edge for production deployments requiring composability, reusability, and fine-grained control — areas where the SDK-first approaches trade away customization for simplicity.
📺 Source: Cole Medin · Published July 10, 2026
🏷️ Format: Tutorial Demo
