OpenAI says computer-use agents near superhuman speed; control becomes the constraint

1. Reliability at the edges — ▶ 1:28

   Weinstein describes the new Astra model's gains as incremental but consequential: it uses fewer thinking tokens, so it is faster on average, and it is better at creative tasks such as drawing in Microsoft Paint. His main point is trust. The gap between a model that is 99% reliable and one that is 97% reliable is hard to perceive but decisive. Sun adds that the system card shows better resistance to prompt injection and fewer unintended actions, which matters for sensitive work like procurement and customer communication.

2. Enterprise is ahead of consumer — ▶ 3:45

   Sun frames consumer as where OpenAI pushes the frontier on new experiences, and enterprise as where capability already exceeds adoption. Large companies run legacy software, 15 to 20 years old, with no MCP tools and interfaces that are hard even for people to operate. Some firms have built skills that teach the model to use these systems. In Sun's telling, enterprises are catching up to existing capability, while consumer use cases are still being discovered.

3. Proactive agents catch dropped balls — ▶ 6:03

   Sun uses Dots, an agent product where each instance gets its own cloud computer, for work that is not the same shape each time. It flags external meeting invites that never reach his calendar. It watches Slack for critical feedback from his reporting chain using a high-level goal, where earlier hourly automations produced notifications he tuned out. It can also log into Sentry through a browser, pull logs, and analyze them, replacing a multi-step manual loop.

4. Speed flips the default — ▶ 8:15

   Weinstein says OpenAI's computer use is now faster than the average human and close to faster than an expert, which he calls a sea change. His reasoning is that background execution tolerates slowness, but anything watched or blocking does not, since users resent feeling they could have done it themselves. Once agents beat experts, he expects the default to be delegating anything doable manually. OpenAI also claims the fastest, cheapest and most reliable product, while offering multiple models at different price points, and cheaper models are improving.

5. Control layers for enterprises — ▶ 10:01

   When agents move faster than humans can audit, enterprises need dials. Sun describes a Guardian or auto-approve model that reviews every action for safety and alignment. Enterprises can also set deterministic sandboxes limiting which apps and websites the agent can use, and require human approval for uploads, downloads, or new sites. Sun argues deterministic controls reassure enterprises more than alignment claims, even for the most aligned model yet. Weinstein adds a visible, animated cursor and permission prompts for new applications to build consumer trust.

6. Agents versus ad economics — ▶ 11:35

   On the open web, Sun identifies a conflict with advertising-funded business models: an agent's ad impression generally does not count as real revenue, so publishers have little reason to welcome agents. OpenAI is working through this with its partnerships team. For sites that do want agents, Sun points to WebMCP, a standard OpenAI supports in its app browser, in which agent and person use the site together. The agent handles tedious steps, and the human browses and then hands back for checkout. No resolution to the ad problem is offered.

7. Universality removes the connector — ▶ 15:34

   Weinstein argues the distinctive property of computer use is that it works with every application by operating it as a human does. Other agent capabilities require developers to build something, such as a connector MCP server. Computer use needs no developer work, so the agent inherits whatever the user can do. Sun adds that complex tasks tend to touch computer use at some point, so he treats it as the model's universal fallback and skips it only for tasks like clicking two buttons.

8. Boring workflows first — ▶ 20:59

   Sun sorts enterprise targets into tedious workflows and ambitious creative work. The first category covers systems of record such as ERP and CRM. Examples are procurement, where a thousand-row spreadsheet passes through ops and procurement teams before vendors are contacted, and month-end finance reconciliation across ledgers lacking APIs. Sales prospecting into CRMs and market research also qualify. The second category is circuit design, architecture, hardware engineering and game design, which Sun expects usage to reach once enterprises understand the model's visual strengths. Enterprises currently focus on the boring opportunities.

9. Cheap software must be better — ▶ 25:36

   Weinstein expects legacy software and new software to coexist for a long time, with computer use pressuring vendors to be agent-friendly. Sun argues that falling software-creation costs raise the quality bar, and that UI is not going away: he built a tailored training tool for his wife's board-game practice, and expects software designed for 100 people or one. Weinstein anticipates traditional interfaces with agents underneath, sometimes using computer use on legacy systems. Sun warns that websites each with their own unconnected agent would fragment the world.

10. Where startups still win — ▶ 33:23

    Sun, a former founder, argues the opening for startups is diffusion rather than capability. OpenAI's teams remain stretched thin despite thousands of employees, and vertical industries adopt AI slowly. He cites a company that photographs restaurant storage, forecasts tomorrow's needs, and uses computer use to order groceries and adjust ERP staffing schedules. For typical businesses, he says, understanding capabilities, connecting tools, deploying, and servicing are separate hurdles, so industry expertise and good UI on top of disparate software remain open opportunities. He names speed as the moat.

11. From power users to billions — ▶ 40:56

    Weinstein says computer use was barely used six to nine months ago, models improved sharply early this year, and the product has grown fast in recent months, though only a portion of users use it. Sun says Atlas's agentic features drew little use, but after folding browser use into existing products he saw YC-level weekly growth. Their next hurdles are cloud agents for phone users and coherence: the Chrome extension, in-app browser, and cloud computer each hold separate cookies and passwords, which confuses consumers.

The strongest claim in this conversation is about speed. Ari Weinstein says OpenAI's computer use is already faster than the average person and approaching the speed of an expert user, the kind he and James Sun represent. If that is right, he argues, the default flips: any task a person could do on a computer is one the agent could do faster, and people would do it themselves only by choice. He calls the transition a sea change in how people interact with computers. The claim comes from the builders of the product and arrives without benchmarks, so it is better read as OpenAI's position than as a measured result.

Weinstein's reasoning shows why speed matters more than it first appears. Agents already have a natural home in background work, where duration hardly counts. But many tasks are watched or block the next step, and there the user's calculation is whether they could have done it faster themselves. Reliability sits alongside speed. Weinstein distinguishes a model that is 99% reliable from one that is 97% reliable, a gap that is hard to perceive but decides whether people trust an agent with real work. His account of the Astra model is accordingly modest: fewer thinking tokens, a little more consistency at the edges, and a lower price tier for less demanding tasks as cheaper models improve.

The structural advantage is universality. Most agent capabilities depend on a developer building something, such as a connector or an MCP server. Computer use works by operating applications the way a human does, so it needs no one's cooperation. Sun explains why enterprises feel this first. Large companies run software that is 15 to 20 years old, with no MCP tools and interfaces that people struggle with. Some have built skills to teach the model to use them. Editorially, the implication is that the agent's reach is set by what a user can do, not by what a vendor chooses to expose.

That reach points the speakers toward unglamorous targets. Sun describes procurement workflows that pass a thousand-row spreadsheet through an ops team and then a procurement team before anyone contacts a vendor. He describes month-end finance reconciliation across ledgers with no APIs, and sales teams moving prospecting data into CRMs. His own example of testing sports-analysis software and feeding feedback back into an app shows the same pattern. Design disciplines such as circuits, architecture and game design come next, he says, but enterprises are not yet thinking about them. The ordering suggests value arrives first where tasks are tedious, repeatable, and trapped behind interfaces nobody wants to automate by hand.

Speed also creates the problem of control. When an agent works faster than anyone can watch, observation stops being a safeguard. Sun describes OpenAI's answer as layered. A Guardian model reviews each action for safety and alignment. Administrators can set deterministic sandboxes limiting which apps and sites the agent touches, and require human approval for uploads, downloads, or new websites. Notably, Sun says deterministic controls give enterprises more comfort than evidence of alignment, even for what he calls the most aligned model yet. That suggests trust in this market will be built through explicit constraints as much as through model quality.

The open web raises a different problem. Sun notes that publishers and advertisers depend on human attention, and an agent's ad impression generally does not count as revenue. OpenAI is working through that question with its partnerships team, but no answer is offered. For sites that want agents, he points to WebMCP, a standard in which the person and the agent use the same site together. The agent does tedious work, the person browses and finalizes a purchase, and the human-facing interface is preserved. The hybrid model is attractive, though it only works for sites that opt in, and the incentive problem for ad-funded sites is untouched.

Both speakers resist the idea that agents make interfaces obsolete. Sun argues that cheap software raises the bar for craftsmanship, and he built a tailored training tool for his wife's board-game practice. Weinstein imagines software designed for 100 people or one. A shared point is that business users still need to understand why a payment was made, which requires a good interface to an audit log. They expect software with ordinary interfaces and agents underneath, sometimes using computer use against legacy systems. Sun warns that websites each running a separate agent that cannot talk to the user's agent would fragment the experience.

Sun's advice to founders follows from the same logic. Capability is no longer the main barrier; diffusion is. OpenAI's teams, he says, still feel stretched thin, and many vertical industries adopt AI slowly. His example is a company that photographs a restaurant's storage, forecasts tomorrow's demand, orders groceries through procurement software, and adjusts staffing in the ERP. For a typical business, he lists four hurdles: understanding what the tools can do, connecting them, deploying, and servicing. These are where industry expertise and good product design can still win. The view comes from a former founder at a lab with obvious interests, but the mechanism he describes is concrete.

Adoption itself looks early. Weinstein says few people used computer products six to nine months ago, models improved markedly early this year, and growth accelerated over recent months, with only a portion of users on it so far. Sun reports YC-level weekly growth after OpenAI folded browser use into existing products, following weak interest in the standalone agentic features of Atlas. The next gates are mundane: coherence across the Chrome extension, in-app browser and cloud computer, which keep separate cookies and passwords, and cloud agents for phone users who never touch a desktop.

The speakers present these as problems of polish and distribution, but a harder one sits underneath. Their control model assumes humans can set limits and review summaries after the fact. Yet the same speed that makes agents attractive removes the possibility of watching them, so oversight depends on sandboxes, an AI reviewing the agent's actions, and logs. Nothing in the conversation shows how well such systems catch the errors a human would have caught. If agents do become superhuman, the audit trail may matter as much as the model.