The New Startup Playbook | Box CEO Aaron Levie

The host asks Aaron Levie how the startup experience has changed since he and his Box co-founders dropped out of college in 2005. Levie says dropping out and starting a company at age 20 was then relatively rare—perhaps a few dozen people were doing it in Silicon Valley, with Facebook helping make the path more visible. Now, he says, dropping out to build a startup has become a standard option for many 20-year-olds. He jokes that he feels “really old,” despite being only a few years older than the host. Levie contrasts the 2005 startup with a 1995 startup: both involved writing code, buying or renting servers, using data centers, and working long hours, so the basic operating model was nearly identical. The 2025 startup, however, looks fundamentally different because of an “AI-first orientation.” A three-, five-, or ten-person team can increasingly produce the output of a 50-, 100-, or 200-person company. He is impressed by the speed of today’s startups, though he also describes some as “crazy and unhinged.”

Levie explains that Box’s early rate of progress was almost entirely determined by how many hours its four founders could spend typing. Living and working in a renovated garage, they typed code, marketing materials, and blog posts; even server management ultimately depended on what the founders could personally do on a computer. Today, he sees founders and early employees increasingly acting as managers of work performed by AI agents rather than as the people who directly produce every artifact. As an example, he describes a founder who claims to review thousands of lines of agent-written code each day. Levie believes reviewing and editing can provide roughly ten times the leverage of creating everything from scratch. One person can launch multiple agents in the background, then review and integrate their output, rather than being limited to one terminal or one integrated development environment at a time. The bottlenecks shift from typing code to reviewing it, orchestrating agents, integrating their work, and correcting mistakes. Levie stresses that agents are not yet a panacea: he compares using them to a slot machine because users must repeatedly try, inspect what comes back, and decide whether it is usable. But if agents can reliably produce a few thousand lines of high-quality code that follows a team’s standards, a startup could gain roughly ten times its previous engineering capacity. That could let it compete on scale and speed in almost any category.

The host points out that cheaper execution makes strategic decisions more important: if it costs little to build something, a company can still move quickly in the wrong direction. Levie agrees that the premium is shifting away from merely hacking together a prototype. Founders now need to write precise specifications, understand the market opportunity, decide what should be built, and demonstrate product taste. A product manager who can clearly specify a product may increasingly be able to build it through agents rather than waiting for an engineering team to implement every detail. Levie says AI agents can neutralize a startup’s traditional disadvantage: a small team may move faster than a large incumbent, but the incumbent could historically deploy many more people and resources. Agents can reduce that resource gap. He believes large companies may have difficulty responding because their workflows, processes, documentation, and engineering methods are deeply embedded in existing systems. Small teams can redesign their operating model around background agents more easily. He expects this advantage to spread beyond coding into marketing, legal work, and outbound sales. Startups that adapt first may have a window in which they can outrun much larger companies in ways that were not previously economically feasible.

The host asks whether companies should rely on one powerful general-purpose AI tool or assemble a collection of specialized agents. Levie clarifies that the central issue is specialization, not necessarily the number of tools. In the Cloud Code-style approach he discusses, teams can create sub-agents responsible for particular parts of a codebase or workflow, reducing the “context rot” that makes a single super-agent unreliable. The best architecture is still unsettled: different startups may choose different tools and methods, and Levie says there is no universal rule yet. He then broadens the opportunity. As AI agents make software possible in areas that previously had little or no software, he expects specialized agents for most domains of knowledge work. Coding is an early example, with companies and products such as Replicate, Cursor, Windsurf, Lovable, and Cognition demonstrating a market that would have seemed implausibly large a decade earlier. Levie says he is comfortable underwriting several hundred billion dollars of market capitalization in AI coding companies that are not incumbents because the market is still expanding. He applies the same logic to life sciences, healthcare, legal work, consulting, and other fields. He imagines a matrix containing every industry and every job function—perhaps 5,000 cells—with opportunities such as sales agents for life sciences or legal-review agents for financial services. Some companies will need to address a horizontal function across industries; others will focus deeply on one vertical.

Levie tells founders to look for industries they understand deeply and build agents for work they have personally performed. He cites Jared Friedman’s advice to do the job an agent is meant to automate, because firsthand knowledge makes it easier to design a capable system. His preferred founding pattern is someone who has already done the work, followed by a team with a co-founder who knows it, and then a founder willing to shadow an expert and learn the workflow directly. He predicts many failures because teams will miss product-market fit, choose markets that are too small, or lack the right expertise. But he also expects surprising successes where the apparent demand for a task is limited only by its price. For example, if high-end legal review costs $1,000 or $2,000 per hour but an agent can perform comparable work for $5 or $10 per hour, companies may suddenly use legal review far more often. More people could file patents, and businesses could fund functions that previously remained at the bottom of their budget priorities. The host jokes that the industry may suddenly produce a surplus of young people who have worked as administrators or in FDA-regulated back offices. He suggests pairing technically capable agent-builders with domain experts recruited from companies such as Pfizer. Levie agrees that a matchmaking model between people who know how to build agents and people who understand specialized workflows could produce many new companies.

The host asks whether consumer startup experience remains a useful proving ground, especially now that AI is pushing many companies toward enterprise products. Levie says there will still be consumer opportunities, but founders should not pre-plan a pivot from consumer to enterprise; Box’s own move from consumer to enterprise was not planned in advance. He is personally more optimistic about enterprise AI because productivity improvements tend to show up more directly in businesses, whose markets are much larger. To explain the difference, Levie uses a point attributed to Anthropic’s Dario Amodei: improving an AI model from the level of an undergraduate chemistry student to a graduate student may not matter much to a consumer asking everyday questions, but it could matter enormously to Pfizer or another life-sciences company. Levie says many consumer use cases are already close to “fully solved” for him. His mother’s everyday questions do not require successive generations of chemistry capability, and in his own family life, AI can already make songs, draw pictures, and create stories that are good enough for his six-year-old. Enterprise applications have much more room to improve. A law firm may still find that an AI system makes 70 mistakes in a contract review, while a higher-capability model could produce meaningful economic value. Levie therefore expects much of the next wave of AI progress to accrue to enterprise use cases, while consumer opportunities will depend more on new packaging and experiences than on simply making the underlying model smarter.

The host asks whether AI creates a new distribution model or merely new product opportunities. Levie gives a deliberately cautious answer: he does not think distribution will change fundamentally over the next five years. There may be new channels, such as an OpenAI app or agent store, but consumer companies will still need to become viral, enterprise companies will still need a go-to-market machine, and companies with technically capable customers may combine product-led growth with sales. His warning is blunt: “If you build it, they will not come.” An impressive agent does not solve the company’s commercial problem, because thousands of startups and competing priorities will be competing for customers’ attention. A viral video or a burst of likes is only a marketing stunt, not durable distribution. Levie defines distribution as owning an audience—a passionate community or cohort that a company can continue reaching through channels such as LinkedIn, X, Reddit, sales, marketing, and public relations. He argues that companies that ignore this work will be overtaken by competitors with better access to customers. The episode closes with Levie emphasizing that AI changes execution leverage, but it does not eliminate the need to identify a market, understand users, build a go-to-market process, and maintain an ongoing relationship with the audience.