OpenAI's Agents SDK v0.13 quietly admits you shouldn't have to run OpenAI models

Version bumps are usually boring. This one isn't, because of what it concedes. The OpenAI Agents SDK hit v0.13 with an any-LLM adapter, breaking the assumption that you'd be running OpenAI models inside OpenAI's framework. Requesty's SDK comparison has the details: alongside the adapter, v0.13 adds opt-in retry policies, MCP resource support, session persistence, and a default Realtime model upgraded to gpt-realtime-1.5.

The any-LLM adapter and the MCP resources are the real story. For a couple of years the implicit deal with a vendor's agent SDK was that it nudged you toward that vendor's models and that vendor's way of wiring tools. LangGraph and CrewAI competed partly by being model-agnostic and protocol-open. OpenAI shipping a first-party any-LLM adapter is OpenAI conceding that ground. They'd rather you build agents on their SDK running someone else's model than not build on their SDK at all.

That fits the through-line of this entire issue. GLM-5.2 says capability is going open and downloadable. The local-first coding agents say own your stack and swap providers. And now even OpenAI's own framework says fine, plug in whatever model you want and keep your MCP tools portable. Model portability stopped being a feature the scrappy open-source frameworks used to differentiate. It became table stakes that the largest lab now has to match. Lock-in is eroding from both ends.

What builders should do: if you're already on the OpenAI Agents SDK, the MCP resource support and session persistence are immediately useful, and the opt-in retries are the kind of thing you'd otherwise hand-roll badly. Voice-agent builders get gpt-realtime-1.5 as the default for free, which is the cheapest upgrade you'll get all month. The bigger move is architectural. Build your agent so the model is a swappable component behind the any-LLM adapter, not a hardcoded assumption. The cost of switching providers is dropping fast, and the labs are the ones dropping it. Design to take advantage of that, because the next price war or export shutdown will reward whoever can repoint at a new model in an afternoon.