What Changed · Source-verified
OpenAI Agents SDK and Agents API: what changed in the stack boundary
The Agents SDK remains an application library; the newer Agents API adds a managed runtime boundary. This expands deployment choice, not the correctness of an agent's policy or output.
Verified change record
- OpenAI maintains the Agents SDK with agents, tools, handoffs, guardrails, sessions, and tracing.
- OpenAI describes the Agents API as a managed agent harness in public beta.
- The API supports managed, adopter-owned, and partner environment choices.
Before
The SDK supplied orchestration primitives while the application owned the runtime and surrounding operations.
After
The Agents API offers a managed harness and environment choices, creating a second deployment path around the SDK concepts.
What did not change
Tool authorization, business acceptance criteria, data governance, and external side-effect safety remain application responsibilities.
Decision detail
- Keep the SDK-only approach when you need direct runtime control and can own operations.
- Evaluate the managed API when long-running execution and sandbox management remove meaningful platform work.
- Do not migrate until data boundaries, evaluation, and rollback are defined for one representative workflow.
- Treat public beta status as a reason to preserve reversibility, not as a reason to reject a bounded pilot.
Evidence
These claims are source-verified. We do not label this page hands-on or benchmarked because no reproducible test artifact is attached.
- OpenAI Agents SDK for PythonOpenAI · checked 2026-09-13
- OpenAI Agents SDK tracingOpenAI · checked 2026-09-13
- OpenAI: Introducing the Agents APIOpenAI · checked 2026-09-13