The Frontier

Ari Weinstein says Astra makes computer use faster by spending fewer thinking tokens, improves creative tasks such as drawing in Microsoft Paint, and feels more reliable than earlier models. He used the gap between 99% and 97% reliability to illustrate why small improvements matter, not as a benchmark result.

James Sun says Astra is OpenAI's most aligned model and is more resistant to prompt injection and unintended actions, which matters when agents handle sensitive enterprise work. OpenAI's Guardian system checks each computer-use action, while administrators can restrict apps and websites or require approval for actions such as uploads and downloads.

Ari Weinstein says computer-use agents are already faster than the average person and are approaching expert human speed. OpenAI sees speed as essential when users must wait for a task to finish, and expects cheaper models to make different quality and price points useful for different workloads.

Ari Weinstein says computer use works across applications without developers building connectors or MCP servers, because the agent operates software through the same interface a person uses. James Sun adds that complex tasks often combine computer use with other capabilities, making it a general interface rather than a standalone task type.

James Sun sees legacy enterprise software as an immediate opportunity because many systems lack APIs or agent tools and remain difficult for employees to use. Companies can build skills that teach agents how to operate these applications, including software that has been in use for 15 or 20 years.

James Sun describes Dot as a proactive cloud computer that can catch missed calendar invitations and surface important Slack messages. He also uses it to review Sentry logs and investigate product feedback, replacing a brittle notification routine that ran every 30 minutes or every hour.

James Sun says agents can automate procurement from spreadsheets through internal purchasing systems and vendor outreach, reconcile finance ledgers, and handle sales prospecting and CRM updates. Procurement spreadsheets can contain 1,000 rows, while finance teams still reconcile records manually because many financial tools lack APIs.

James Sun also sees computer use helping with market research and creative enterprise work, including design, architecture, hardware engineering, and game design. Ari Weinstein points to consumer uses such as filling out DMV forms, unsubscribing from email, managing personal spreadsheets, and making custom plans for clients.

James Sun says agents can connect to websites but raise a revenue problem for publishers and advertisers when agent activity does not count as a paid ad impression. OpenAI supports WebMCP in its app browser as one way for people and agents to use a site's interface together rather than replacing it with an API.

Ari Weinstein expects legacy software and new AI-native products to coexist: computer use can operate old systems, while cheaper software creation makes new products easier to build. James Sun argues that carefully crafted interfaces will remain valuable, especially for personalized tools that serve a person or a small group.

James Sun argues that enterprise agents still need a usable interface for setting goals, reviewing decisions, and tracing actions such as payments. Ari Weinstein says product design should match the workflow: foreground collaboration for individual users, or background execution with concise updates that avoid overwhelming people.

James Sun says startups can still win by pairing AI tools with deep industry knowledge, regulatory understanding, customer service, and go-to-market skill. He sees room in industries where AI adoption is slow, especially when agents can connect complex workflows across separate systems.

James Sun describes a restaurant workflow that uses storage photos and projected traffic to estimate missing supplies, place grocery orders through procurement software, and adjust employee schedules in an ERP. He argues that businesses still need help selecting, connecting, deploying, and servicing AI tools.

Ari Weinstein expects computer use to move from background task execution toward real-time collaboration, voice interaction, and multimodal input and output. James Sun says the product also needs a more coherent browser experience because separate local, app, and cloud browsers can have different passwords, cookies, and tabs.

Ari Weinstein says computer-use models became strong enough for useful products earlier in the year, after which OpenAI teams built product experiences around them. He describes the work as accelerating over the past six to nine months; James Sun says browser use has since grown rapidly week over week, with cloud access as the next adoption step.

Aaron Levie says dropping out to start a company was unusual when he left college with his co-founders in 2005, but has become a standard path for young founders. He sees today’s AI-first startups as a bigger operating shift than any earlier period in his career.

AI agents let small teams produce far more work than their headcount once allowed. Levie says a three-to-ten-person startup can approach the output of a company many times its size, giving founders leverage that used to require large teams.

Founders are shifting from producing every deliverable to managing agents that write code and create marketing assets. Levie says one founder told him he reviews thousands of lines of agent-written code daily, making review and orchestration the new bottlenecks.

Agents can narrow the resource gap between startups and incumbents, Levie argues, because a small company can run many agents in parallel while large firms remain tied to established processes. Agents still make mistakes, so the advantage depends on reliable output and effective review.

As coding becomes cheaper, Levie says clear specifications, market judgment, and product taste become more valuable. Founders who know what to build and can define it precisely give agents direction that faster execution alone cannot provide.

Levie describes specialization as a way to limit AI context rot: agents can own parts of a codebase or workflow, including through subagents. He says the right architecture varies by team and tool, with no settled best practice yet.

Levie expects startups to build agents for knowledge-work roles across industries, including life sciences, healthcare, legal work, and consulting. He cites coding agents as an early market, naming Replit, Cursor, Windsurf, Lovable, and Cognition, and sees substantial room beyond incumbents.

Cheap agents could unlock work that businesses currently leave undone because human labor costs too much. Levie says a steep price drop could reveal far more demand for services such as legal review, while low-cost agents could justify smaller, previously unfunded tasks.

Levie advises founders to understand the job their agent will automate. He cites Jared Friedman’s advice to do the work firsthand, and says founders can bring that expertise themselves, recruit a domain-expert co-founder, or closely observe practitioners.

Levie says the founding team matters more when agents multiply each founder’s output. He would look for a market he understands deeply and reverse-engineer its work, while noting that some agent opportunities will fail because the market is too small or product-market fit is weak.

Levie expects enterprise AI to capture much of the value from model improvements because higher accuracy matters directly in work such as drug development and legal review. He sees consumer opportunities too, but says many personal uses already meet users’ needs and future gains may come from better app packaging.

Levie does not expect AI to remake go-to-market fundamentals. Consumer startups still need virality, enterprise startups need a sales engine, and product-led growth can complement enterprise sales when customers can adopt through self-service.

An impressive agent is not enough to win customers, Levie argues. Startups need sustained access to a clearly defined audience through channels such as LinkedIn, X, or Reddit, backed by sales, marketing, and PR rather than a one-off viral stunt.

Levie says advanced startups now describe AI as a potential three-to-tenfold source of leverage, beyond the earlier use of autocomplete and coding assistants. Their small-team advantage comes partly from avoiding the documentation, tooling, and process changes larger organizations need to run agents at scale; he is studying how to bring those lessons to Box.

Ben Horowitz says a great entrepreneur feels compelled to build a company because of an idea. He tells founders to expect self-doubt and judge themselves by what they do, not by whether they feel confident.