Summary
Computer-using agents are now faster than the average person and approaching superhuman performance, according to OpenAI’s Ari Weinstein and James Sun. In this a16z podcast, hosts Sema Amble and Fbritzio Saraphiniti talk with the two OpenAI team members about how these agents are moving into real work inside companies.
The guests explain what OpenAI’s newest model adds to computer use: fewer thinking tokens for faster results, stronger creative performance such as drawing in Microsoft Paint, and better reliability, which they describe as the gap between 97% and 99%. They also discuss improved resistance to prompt injection and fewer unintended actions, which matter for sensitive enterprise work like procurement and vendor communication.
The conversation then turns to enterprise adoption. OpenAI describes two layers of trust: deterministic controls and demonstrated alignment. Admins can sandbox an agent to specific apps and websites and require human approval for downloads, uploads, or visits to new sites. The guests also cover consumer experiences, design choices like a visible animated cursor, and why computer use works across software without the custom connectors or MCP servers that other integrations need. The episode closes with advice on which tasks companies should consider automating first. It is a useful look at where computer-use agents stand and how organizations can deploy them safely.
📺 Source: a16z · Published October 26, 2026
🏷️ Format: Podcast
