Three separate companies shipped something this week that, read together, says more than any one of them does on its own. OpenAI put its Codex agent harness behind a public API. Cursor launched a mode called Projects that runs on its own cloud machine and keeps going after you close your laptop. Microsoft quietly let Visual Studio developers plug in whatever model they want instead of whatever ships by default. None of these are really about a smarter model. They're about giving up on the idea that a person is watching.
Start with OpenAI's Agents API, out in public beta on September 10. It exposes the same session orchestration, context compaction, and multi-agent coordination that powers Codex as a plain API call, and it can run inside an OpenAI-hosted sandbox instead of your own infrastructure. The sandbox billing isn't the interesting part; containers cost what containers cost. What matters is that a "subagent" is now a first-class primitive. A parent agent can split a task, hand pieces to other agents running in parallel with their own separate context, and stitch the results back together when they're done. That's not autocomplete. That's a small org chart made of API calls.
Cursor's Projects, announced the same day, is the same idea from the editor's side. A coordinator agent doesn't write code itself. It plans the work, delegates to agents that implement it, and checks the results before handing them to you. It's built explicitly to run "over months," including recurring work you never have to re-prompt for, on a machine that keeps executing after you've stopped looking at it. Cursor is betting that developers want to assign a feature the way you'd assign it to a contractor, not babysit it turn by turn in a chat window.
Here's the part worth sitting with. None of this changes what these systems get wrong. It just changes how long it takes you to notice. A bad autocomplete suggestion costs five seconds. A subagent that misread the task and has quietly been building on a wrong assumption for three days, inside a project you started and then forgot about, costs you the three days plus however long it takes to untangle what it built in the meantime. The tooling for "come back and see what changed" is nowhere near as mature as the tooling for "go do this." If you use either of these, decide what checkpoint you expect before you walk away, because the product itself won't remind you to check.
The quieter story is Visual Studio's September update, which put Bring Your Own Key into the IDE by default across every SKU, Community included, starting September 8. Point Visual Studio's Agent mode at a Microsoft Foundry deployment, or at another provider's key, and Copilot stops being the only door into AI-assisted coding in that editor. That's a real admission from Microsoft: however good Copilot's underlying model is this month, they don't expect you to want to be locked to it forever. Pinning a whole workflow to one vendor's agent stack has never looked riskier than it does the same week three different companies shipped a way around doing exactly that.
If there's one practical takeaway from a week like this, it's to treat every "runs without you" feature as an experiment with an unknown blast radius until you've watched it fail once on something that doesn't matter. Let a Project or a subagent chew on a low-stakes migration first. Find out what it does when it's wrong before you find out what it does when nobody was looking.