Sep 30, 2026 · 39 min
OpenAI pushes computer use toward an AI-native cloud
Why Dwarkesh is Wrong about Computer Use + How OpenAI shipped its Jev competitor in 1 Week
The episode connects more capable computer-use agents with the infrastructure needed to make them fast, reliable, and useful across real software.
- 1Computer-use models are improving through debugging, retries, richer interface data, and code-generating harnesses.
- 2OpenAI’s agent stack is shifting toward asynchronous tools, bidirectional connections, faster inference, and compacted context.
- 3The platform strategy aims to expose higher-level primitives for long-running agents, much like foundational services in cloud computing.
Don't miss
Ari Weinstein’s account of agents testing the software they build turns computer use from interface automation into an engineering feedback loop.
The brief
Ari Weinstein argues that computer use has not stagnated: models now debug, retry, and introspect, while screenshots, accessibility data, and the DOM give them richer interfaces.
Persistent Linux computers let agents handle websites and workflows without APIs, from customer service and shopping to software testing and visual playtesting.
The conversation’s tension is speed versus trust: agents may become expert or superhuman, but timing, safety checks, and high-stakes delegation remain unresolved.
Nikunj Handa explains how OpenAI’s stack is being rebuilt around async tool calls, WebSockets, UltraFast inference, caching, and context compaction.
The episode ends with a larger claim: OpenAI is assembling an AI-native cloud, where developers use higher-level primitives to build responsive, long-running agents.