OpenAI’s Jev clone could help the frontier lab stop its swarming agents

Sam Altman introduced a Decisions API at DevDay that gives Luna a predefined set of options — classifications or agent behaviors — for extremely fast choices while keeping multimodal and safety features. TypeSafe AI’s Jev, from former OpenAI engineer Diogo Almeida, popularized the cheap, calibrated classifier pattern weeks earlier; Almeida joked about “clone wars” on X.

TechCrunch notes Decisions API is in limited preview and that OpenAI has been using a separate watcher model at significant compute cost after agent incidents. QueryStory’s Shapor Naghibzadeh suggests Jev-like models could make that monitoring affordable. Calibration quality, not just speed, is the open question; Almeida cites synthetic data as TypeSafe’s moat.

We’re interested because agent security dies when the babysitter costs more than the baby.

Constrained decision models are useful AI in the boring sense: fast gates, explicit option sets, measurable calibration. If OpenAI can run a Decisions-class watcher on every sensitive agent step, swarming failures get cheaper to catch. Cloning the startup pattern overnight is rough on TypeSafe; it’s also how platforms absorb good ideas.

I’m watching real calibration numbers in the preview, not demo latency. Dice are fast too.

We’ll root for monitoring that actually ships in the agent loop. Speed without calibrated refusal is just a faster wrong turn.

TechCrunch, September 30, 2026, on OpenAI’s Decisions API and Jev.