A real pydantic-ai v2.36.0 typed Agent runs in-process with the Mission SDK, Closure Resolution and governance — a genuine typed agent invoked with run_sync, its validated model projected at the Mission boundary. No mock, no sidecar approximation.
PydanticAI's typed boundary guarantees an output's structure — a well-typed, schema-valid model. Structure alone does not establish evidence support, freshness, authority, or policy admissibility — nor exactly-once execution across a restart. ReDevOps supplies those runtime checks as shared execution infrastructure: verification, governance and replay, orthogonal to validity.
Four schema-valid Pydantic outputs the runtime still refuses to act on — unsupported by evidence, stale, no authority, governance-denied. One is admissible and passes both layers. The two guarantees are provably distinct.
A real typed intent (rebalance TLT) wrapped in a Mission. Approval held; exactly-once execution; replay reconstructs the typed state to an identical hash and fires zero extra side effects.
The same typed verdict agent on naive vs Closure-Resolution context, equal model & budget. Evidence recall 0.841 → 0.877; selective autonomy 0.889 @ 90% coverage.
PA-D is deterministic and decisive — the two layers are provably orthogonal. The SciFact closure gain is modest because the corpus is content-saturated; it is the same closure mechanism the NVIDIA slice measures on a different framework agent. That cross-framework agreement is the conformance point, not a ranking.
// pydantic-ai 2.36.0 · Mission SDK 0.2.0 · Context Runtime 0.2.0 · one Python env · github.com/redevops-io/redevops-integrations