OpenAI's Agents API won't kill agent frameworks

Every time OpenAI ships something for agents, I see people saying “frameworks are dead.” After actually running agents in production, I’m not sure I buy that. The Agents API is pretty good, but I think there’s a point where you still want to own the orchestration layer. For me, it comes down to a few things: Models: In a real system, I’m probably not using one model for everything. Maybe a small model for routing, a vision model for parsing, and a bigger model for the harder reasoning. I’d rather be able to swap those without changing the whole architecture. State: I don’t always want session state, tool calls, and execution history living somewhere outside my infra. Sometimes that stuff needs to be in my own DB/VPC. Failures: This is probably the biggest one. If an agent gets to step 9 out of 10 and dies, I don’t want to start from step 1 again. I want the execution history so I can resume, replay, or pause for human approval before a side-effecting action. Evals: Production failures are basically free eval data. If I own the traces/journal, I can turn those failures into regression cases and test new prompts/models against them. So I don’t think frameworks are automatically dead just because the model providers are adding agent APIs. Thin wrappers around an LLM loop? Sure, those are getting commoditized. But durable execution, state, replay, human-in-the-loop, and model portability are a different problem. Curious where others draw the line. What would make you comfortable keeping orchestration/state entirely on a vendor API, and when would you want to own it?