OpenAI: When AI Agents Use Computers Faster Than You

Computer-using agents are already faster than the average person at some tasks and are approaching superhuman performance, according to the episode’s introduction. The software around a model—the harness—can potentially make it nearly twice as fast. OpenAI describes its computer-use product as the fastest, cheapest, and most reliable, and says less expensive models are improving. Ari Weinstein says Astra feels somewhat faster in his experience because it uses fewer thinking tokens for the same task, and he finds it more capable on some creative work and more reliable at the edges. James Sun calls Astra OpenAI’s most aligned model, citing stronger resistance to prompt injection and fewer unintended actions; he also says it beat GPT-5.6 in OpenAI’s internal video-game tests.

Computer use lets an agent operate software through the same interfaces a person uses, rather than requiring a dedicated API, connector, or MCP server for every application. That makes it relevant to legacy enterprise systems that are difficult to use or lack MCP tools; companies can teach agents how to navigate them through skills. In Dots, an agent can use a cloud computer and full browser for tasks such as buying items or booking appointments. Sun says it noticed an outside meeting invitation missing from his calendar and offered to accept it, and he uses it to flag Slack messages tied to goals such as tracking feedback from his reporting chain or surfacing critical product feedback.

Cross-application work is a major enterprise opportunity, especially where information is scattered across systems. An agent could take data from team spreadsheets into procurement software and then contact vendors, review ledgers and reconcile records in finance systems, or move sales information from transcripts and data tools into a CRM and outreach software. Agents could also conduct market research by creating accounts in competing products, testing them, capturing screenshots, and reporting observations. Sun describes a restaurant workflow that could use photos of storage and traffic projections to identify grocery needs, place orders through procurement software, and adjust employee schedules in an ERP.

Speed affects whether these tools feel useful even when tasks run in the background: people may wait for a result or watch an agent to verify it, but if it takes too long, doing the task themselves can seem easier. Sun says Astra can ask for clarification when a request is vague, and may proceed if the user does not respond in time. OpenAI’s enterprise controls can limit which apps, websites, or data sources an agent accesses and require human approval for actions such as downloads, uploads, or visiting new websites. Guardian checks actions for safety and consistency with the original intent; permission prompts and a visible cursor make activity easier to monitor. The speakers stress that alignment and evaluations can help build trust, but deterministic restrictions and human approval offer additional safeguards amid questions about oversight and auditability.

Agents could unsettle ad-funded websites if their activity does not count as a real ad impression. OpenAI says its browser supports WebMCP, allowing people and agents to use a website together while keeping the site’s interface involved instead of replacing it with an API. Still, people may want to browse, compare options, or complete transactions themselves. The speakers expect traditional and AI-native software to coexist: Sun does not expect every workflow to move to MCP or APIs, or computer use to become obsolete. He values crafted software and direct interaction, while anticipating that cheaper software creation will make tailored tools for niche needs more practical. A future product could combine a purpose-built interface with an agent operating older software behind it.

OpenAI’s computer-use work has been underway for about six to nine months. The models improved earlier this year, and usage has grown rapidly in recent months, though only some users use the tools. Atlas’s agentic experience initially attracted few users, partly because people treated it like an ordinary browser; after model capabilities improved and browser use was incorporated into products with agentic features, growth was described as “YC level” week after week. Making extensions, in-app browsers, and cloud computers feel coherent despite separate cookies, passwords, and browser states is a focus for the next quarter. Ari expects agents to get faster, combine interfaces with multimodal inputs and outputs, and collaborate with people more in real time; bringing computer use to phones could broaden access beyond today’s largely power-user desktop experience. Sun says startups still have room to build, particularly where domain expertise, regulatory knowledge, customer service, or complex industry workflows matter.