OpenAI: When AI Agents Use Computers Faster Than You

Seema Amble opens by asking Ari Weinstein and James Sun what has moved computer-using agents forward, especially with OpenAI’s Astra model. Ari says computer-use models were already strong before Astra—he describes earlier models such as “Soul” as excellent—but Astra improves several important edges. It is faster on average because it uses fewer thinking tokens for the same task, performs better on creative visual work such as drawing in Microsoft Paint, and feels more reliable. Ari emphasizes that the practical difference between 97% and 99% reliability is significant when an agent is taking actions on a user’s behalf. James adds that Astra is OpenAI’s most aligned model and is substantially better at resisting prompt injection and avoiding unintended actions. That matters for enterprise use cases involving procurement, vendors, and customers. He also says Astra performs strikingly well in video games: OpenAI tested it against “5.6,” and Astra won every time; people have posted examples of it beating Fallout in roughly 30 hours. The guests frame these improvements as relevant not only to entertainment but also to the model’s ability to navigate complex visual interfaces.

The hosts ask how consumer and enterprise computer use differ. James says consumer products are where OpenAI pushes the frontier of what a faster, better model can make possible, while enterprise use is already broad because companies operate large amounts of old software without modern APIs or MCP tools. Many systems have been in place for 15 or 20 years, with interfaces that are difficult even for people to operate. Companies are responding by building “skills” that teach agents how to use these systems. Ari describes consumer use as a way to remove tedious work from daily life. In the “dots” product, each dot can have a computer in the cloud and a full web browser, allowing it to buy items, book a haircut, or handle other online tasks. James says dots are especially useful for proactive, long-running work rather than fixed workflows. For example, dots can find external meeting invitations buried in email that do not appear on his calendar, ask whether he wants to accept them, and monitor Slack for important messages from people in his reporting chain. When a relevant issue appears, a dot can use a cloud computer to inspect Sentry logs, analyze them, and report back, replacing a chain of manual searches, downloads, and Codex prompts.

Ari says OpenAI has focused heavily on speed, cost, and reliability and believes its computer-use product is now the fastest, cheapest, and most reliable available to them. Background tasks can take time without disrupting the user, but speed matters when a person is waiting, watching, or deciding whether it would have been easier to do the work manually. He says agents are already faster than average humans at using computers and are approaching a “superhuman” point where they will outperform expert computer users. Once that happens, any task a person can perform manually could usually be delegated by default, unless the user specifically wants to do it themselves. The guests also discuss the tradeoff between speed and oversight. James says that as agents become too fast for people to watch every action, enterprises need control over how much human involvement remains and how actions are audited. OpenAI’s enterprise controls allow administrators to decide whether actions are automatically approved or require intervention. Its Guardian system reviews actions to judge whether they are safe and aligned with the user’s intent. A consumer may prefer an agent that simply completes everything, while an enterprise might require approval for certain actions. The broader point is that autonomy, safety, reliability, and speed must become adjustable settings rather than one fixed operating mode.

The hosts ask how software companies are responding to agents that can operate legacy workflows. James says the situation is relatively straightforward inside company-approved software: businesses generally want employees to automate tedious work. The open web is more complicated because some sites rely on advertising or publishing revenue. If an agent uses a site on a person’s behalf, it is unclear whether that activity generates a meaningful ad impression or other revenue, so OpenAI is working through the issue with its partnerships team. For companies that want agents to use their sites, James points to WebMCP, a standard OpenAI supports in its browser. WebMCP could let people and agents use a site together rather than forcing the agent to replace the entire human experience. He gives the example of shopping or booking travel: an agent might handle tedious searching and preparation, while the person continues browsing hotels or flights and takes over for the final transaction, including entering a credit card. The hosts then discuss whether computer use will preserve old enterprise software or be used to build entirely new applications. Ari expects both outcomes. Decades-old software will not disappear overnight, but the fact that agents can use it may pressure vendors to make their products more agent-friendly. At the same time, cheaper software development will produce new AI-native applications.

James divides enterprise opportunities into two broad categories: automating tedious workflows and enabling more ambitious creative work. The first category is already practical across ERPs, CRMs, and other systems of record. In procurement, for example, teams may fill out a spreadsheet with hundreds or thousands of rows of needed lab equipment; operations then transfers it into procurement software, and the procurement team contacts vendors. A computer-using agent can connect those steps even when the underlying systems lack APIs. Finance is another strong use case. Companies maintain multiple ledgers for expenses and other activities, and month-end reconciliation often requires people to move among financial systems that cannot easily expose their data. An agent can inspect the ledgers, reconcile information, flag anomalies, and investigate further. Sales teams can similarly use agents to conduct prospecting, gather information from transcript and data tools, and enter the results into a CRM or outreach system. James also describes using computer use for market research in a personal badminton-video project: the agent created accounts on software for pickleball, squash, and tennis, tested the products, captured screenshots, and returned feedback. Ari adds consumer examples such as completing DMV forms and unsubscribing from unwanted email. He also uses agents for personal finance and meal-planning spreadsheets, while a personal-trainer friend uses them to create customized plans. Beyond routine work, the guests expect agents to support visual design, circuit design, architecture, hardware engineering, and game design.

The guests reject the idea that software will become only a set of invisible agents talking to one another. James says craftsmanship still matters: a well-designed product understands how a particular user wants to work. He describes building his wife a tactical trainer for a board game in which players connect five pieces to win—apparently Gomoku or a similar game—because he knew how she preferred to learn. An agent could summarize her mistakes afterward, but the interactive training experience itself benefits from deliberate human-designed software. Ari makes a similar case for direct manipulation. People are hardwired to move and control things with their hands, and existing software is often designed for millions or billions of users. Agents could make it practical to build software for a hundred people or even one person, producing more personal interfaces. Seema adds that enterprise users may want agents to perform the work but still need an enjoyable interface for setting goals, reviewing logic, and examining an audit log—for example, understanding why a payment was made and what discount was applied. Looking ahead, Ari expects software that combines traditional interfaces with agents operating legacy systems underneath. James says the challenge is making the experience coherent across ChatGPT, browser extensions, desktop apps, and cloud computers, which currently have different cookies, passwords, and browser states. He and Ari are especially excited about real-time collaboration, multimodal interaction, and agents that proactively come to users with useful work instead of requiring users to remember to invoke them. The agents should adapt to people, not force people to adapt to the agents.

When asked what advice they would give a startup using computer use in an enterprise workflow, Ari says the right human-agent relationship depends on the product’s modality. A person using software more efficiently may need a collaborative interface, while a cloud-based system operating other applications may need different rules for escalation, error reporting, and summaries. Founders must decide when humans should approve actions, how agents should report failures and successes, and how to avoid overwhelming users with a notification for every event. James says startups still have substantial room despite the capabilities of the major AI labs. OpenAI itself has thousands of employees but remains stretched across many problems. A startup with deep industry knowledge, regulatory expertise, customer relationships, and a clear go-to-market process can use the available models to move faster and solve workflows that large labs will not package themselves. He says speed and pain—solving a real, costly problem quickly—are important sources of advantage. He gives the example of a company serving restaurant groups. The system uses photographs of back-room storage, vision models, and projected next-day traffic to determine what supplies will be needed. Computer use then orders groceries through procurement software and adjusts employee schedules in an ERP based on expected demand. Restaurant groups may not know how to assemble these tools themselves, so there is an opportunity to connect disparate software, deliver a complete end-to-end workflow, and present it in a UI that feels natural to the customer. The guests close by calling this the “scaling moment” for computer use. OpenAI has seen strong growth over the previous six to nine months as browser and agentic capabilities have been integrated into existing products. The next challenge is bringing cloud computers to ordinary phone users and making the experience coherent, understandable, and trustworthy.