OpenAI engineers detail new agent tools like Computer Use and Decisions API

In this podcast episode, two OpenAI product leads discuss the company's expanding agent technology stack following its DevDay event. Ari Weinstein, who leads Computer Use, explains how the tool has improved dramatically in recent months, becoming faster and more reliable at completing real-world software tasks, sometimes outpacing average human speed. He describes new techniques combining screenshots, accessibility data, and code generation to help agents debug and recover from errors.

Nikunj Handa, from OpenAI's API team, outlines new developer features including async tool calling, the Decisions API, prompt caching, and UltraFast inference. The conversation also touches on whether OpenAI is effectively building its own 'AI cloud' with infrastructure for agent memory, computing, and state management.