The New Startup Playbook | Box CEO Aaron Levie

Anthropic

Across the shows: 44 speakers in other episodes, 26 bullish and 44 bearish calls, past 30 days

• Aaron Levie · BoxBullish

  Claude Code's subagents solve context rot; a single super agent doesn't work yet.

  The key point is that a single super agent does not exist and does not work today, so some specialization is needed. Claude Code solves this 'context rot' problem by letting users create subagents that each own part of a codebase or workflow, which is a form of virtual specialization. Specialization can take many forms, and best practices are still emerging, with different teams getting different answers. Levie presents this as the current mechanism that makes agents usable at scale, not as a claim about Anthropic's market position.

  “Claude Code has created a way to have that virtual specialization by just letting you go and create sub agents that own a particular part of your code base or a particular part of your workflow.”at 12:28

Enterprise AI adoption & AI software

• Aaron Levie · Box

  AI agents let tiny startup teams outrun incumbents, which was never economically feasible before.

  AI-first teams of 3-10 people can produce the output of a 50-200 person company, because agents remove the typing bottleneck. Engineering is the first example: founders now act as managers and reviewers of background agents running in parallel, so their speed is no longer limited by how fast one person can write code. Incumbents have many workflows and processes wired into how they operate and cannot change fast enough, which opens a window where startups can outrun them. Levie hedges that agents make mistakes and are not yet a panacea.

  “So you have a window where startups can actually outrun these big companies in a way that was just never economically feasible before, and it's effectively all due to AI agents.”at 8:44

• Aaron Levie · Box

  AI coding startups could reach a couple hundred billion in market cap outside incumbents.

  Replit, Cursor, Windsurf, Lovable and Cognition already show that AI coding can create large startup valuations, which would have sounded implausible a decade ago when GitHub was the only notable exit in the space. Levie says that is only the start, and he is comfortable underwriting a couple hundred billion dollars of market cap in AI coding going to new companies rather than incumbents. His reasoning is the very large addressable market, which must be served.

  “I am perfectly comfortable underwriting a couple $100,000,000,000 of market cap in the AI coding space that don't go to incumbents purely because there's just so much service area, so much TAM that needs to exist.”at 15:26

• Aaron Levie · Box

  AI agents give engineers roughly 3-10x leverage, versus 10-20% gains from earlier coding assistants.

  A year ago, advanced startup founders reported 10-20% productivity gains from tools like GitHub Copilot, mostly autocomplete. The most advanced founders now describe a three-to-ten-times leverage effect because they can run background agents. A founder he talked to reviews thousands of lines of agent-written code a day, and Levie says the reviewer's leverage is about 10x that of the person who writes the code. Existing engineering organizations are not set up for this: codebases lack documentation, tool stacks and methodologies are not ready. Small teams therefore benefit most. Agents are still a slot machine that needs many retries.

  “It's not a 20, 30% form of leverage. It's possibly a three to five to 10x form of leverage because of this ability to have background agents.”at 17:11

• Aaron Levie · Box

  Hundreds of $1B+ AI agent companies will emerge, each automating a job function across industries.

  Levie describes a matrix of every industry against every job function, roughly 5,000 cells, and expects an AI agent company for most of them, such as life sciences sales reps or financial services lawyers. Agent companies will sell to corporations, services firms and system integrators. Some markets will be TAM-constrained and need to be built horizontally or vertically, and many will fail. He expects surprises where demand is larger than expected. Startups will outrun both traditional software incumbents and traditional services firms in these spaces, and incumbents will take only some categories. Agents also bring software to industries that never had it.

  “I really think that there will be hundreds of companies that are all 1,000,000,000 market cap, 5,000,000,000 market cap, 10,000,000,000 market cap and beyond”at 19:59

• Aaron Levie · Box

  Most AI progress will flow into enterprise, since consumer AI needs are already saturated.

  AI is fundamentally a productivity tool, and productivity shows up meaningfully in the enterprise, so Levie expects enterprise to be the larger TAM. Consumer use is already saturated: his mother's use cases are fully solved, and for his own parenting AI could stop now and still be a big net positive. Consumers would not notice a jump from undergraduate to graduate-level chemistry, but a PhD-level jump matters for Pfizer and life sciences. A law firm running a contract through a model can still find about 70 mistakes. Consumer startups can still win, but through new packaging, not better models.

  “So by definition, and back to the original point, by definition, most AI progress is going to now amortize into enterprise use cases.”at 29:08

AI & jobs

• Aaron Levie · Box

  Agents priced at a fraction of human labor would unlock far more demand, as in legal review.

  Labor supply and demand are not in equilibrium. Expensive labor goes unbought: legal review costs $1,000-2,000 an hour, so many firms simply cannot afford it. If an agent can do it for $5-10 an hour, demand could rise about 50x, creating a new category. Levie also describes internal functions that always sit at the bottom of the budget list because they are too expensive. Hiring a person costs $100-300K, while an agent might cost about $1,000, with no ceiling on ROI and no minimum price floor. He expects some markets to be TAM-constrained and many startups to fail.

  “if somebody could have introduced labor at a tenth of the cost that it previously was priced at, that there would be actually 50x the demand for that particular form of labor.”at 22:35

Venture capital & startups

• Aaron Levie · Box

  AI won't fundamentally change distribution; startups still need audiences and go-to-market machines.

  Levie hedges that he may be wrong in five years, but he does not think AI disrupts the distribution model of tech. New channels, such as an OpenAI app store, will appear, but the basics stay the same. Consumer companies must go viral, and enterprise companies must build a go-to-market engine, with product-led growth plus a sales motion where the audience is technical. A great agent does not sell itself: startups get drowned out by thousands of others and need to own a community on LinkedIn, X and Reddit. A viral stunt is not distribution.

  “possibly very wrong in five years from now. I don't think distribution changes that much.”at 30:27