On an episode of Latent Space recorded at OpenAI DevDay, product leaders argued that computer use has fundamentally changed in just a few months. Ari Weinstein, who leads computer use at OpenAI, said agents are now far better at debugging and recovering from failures, and can complete some tasks faster than an average human. The shift comes from fusing multiple input signals—screenshots, accessibility trees, the DOM, Playwright, and generated JavaScript—rather than relying on any single approach.
The same episode detailed how OpenAI shipped its Decisions API in about a week, a move that the hosts note mirrors patterns popularized by the Jev podcast and community. Nikunj Handa from OpenAI's API team described the API as a wrapper around GPT-6 Luna for now, but stressed the team is motivated and egoless enough to clone good ideas. The API supports real-time decision-making, routing, and classification, and is already used internally for support workflows.
Both segments point to a broader theme: OpenAI is pushing toward higher-level primitives—whether that means agents that can operate computers more reliably or APIs that handle decisions directly. The source material is a single podcast episode, so there are no independent disagreements to note, but the framing suggests the pace of change in agentic AI is accelerating faster than public perception.