Capabilities, Not Tools: The Agent Assembly Shift

Pydantic AI 2.0 redefines agent construction around a single primitive: the capability. This bundles instructions, tools, hooks, guardrails, and MCP servers into reusable units. It’s a shift from ad-hoc assemblies to composable building blocks, promising to reduce duplication and enforce consistency across production agents.

The framework now supports progressive disclosure, letting agents carry many capabilities but only load full instructions when needed—cutting token waste. However, this optimization introduces management overhead: teams must curate capability catalogs, version them, and ensure they remain coherent as they evolve.

The lean core vs harness split signals a maturing ecosystem. Critical functions like thinking and web search are built-in; optional features like code execution live in the harness. This design keeps the core lightweight but may complicate dependency tracking for engineering leaders balancing speed with stability.

The speaker claims Pydantic AI is “leading the industry again,” but evidence is thin. No production benchmarks or cost comparisons are provided. The demo is trivial. Engineering teams should evaluate whether this composability model integrates with their existing toolchains before adopting, as lock-in risk is real.