OpenAI and Cursor agree on agent coordinators. They disagree on who runs them.
The New Stack Robert Kimani
OpenAI and Cursor both shipped agent coordinators. They agree on the setup, but not on who should be in charge.
Based on reporting by The New Stack, Robert Kimani — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI opened its Agents API in public beta, and Cursor launched Projects on the same day. Different products, same underlying idea: one coordinator handles the bigger job, then hands pieces to specialized agents that do the actual work.
That split is becoming the default answer to a problem the first wave of coding agents never really solved. A single agent can get through a small task just fine. But once the work turns into a migration, a refactor, or a mess of dependencies, the context gets crowded, compaction starts mangling what matters, and accuracy slips.
Hilliary Lipsig, a senior principal site reliability engineer at Red Hat who leads Azure Red Hat OpenShift SRE teams, says the industry keeps relearning this lesson. She compares multi-agent orchestration to Kubernetes: the coordinator becomes the source of truth, routes work, and recovers from failure states, while workers stay focused on their own jobs. That’s the important shift here. The question is no longer how smart one model can be. It’s how much reliable work a system can complete when several capable agents are involved.
The catch is that splitting the work creates its own problems. More agents mean more tokens, more coordination, and more ways for things to drift apart. Cursor moved its cloud-agent loop to Temporal to get durable execution and retries, and says that setup now handles 50 million actions a day across 7 million unique workflows. OpenAI’s own Agents API supports U.S. data residency, but not Zero Data Retention, and a self-hosted sandbox does not change that.
Security is where the architecture gets real. The coordinator decides who sees what, which tools get used, and when a human needs to step in for something irreversible. That matters because agent systems can leak context, overreach permissions, and produce ugly failure modes when authority is too broad. The industry is treating this like infrastructure now, not a clever demo, which is probably the first sane thing anyone has said about AI coding in a while.
My take — AI-written commentary, not fact-checked reporting
The real product here isn’t the coding agent. It’s the permission system wrapped around it, which is a very unglamorous place for the future to live. Everyone loves autonomous software until it starts asking for root access like a bored intern with a credit card.
Read more about this at: The New Stack
