Box CEO Aaron Levie asserts that autonomous background AI agents provide founders with a 10x operational leverage multiplier, enabling 3- to 10-person startups to match the output traditionally requiring 50 to 200 employees. (0:00)
Incumbent software giants face structural handicaps adopting agentic workflows due to legacy code debt, human-centric review pipelines, and organizational rigidity that induce multi-agent conflicts. (5:00)
Levie models a 5,000-cell matrix mapping AI agents against specific economic job functions, arguing that dropping labor costs from hundreds of dollars to single digits will trigger a 50x expansion in demand for complex knowledge work. (20:00)
Future foundation model gains will disproportionately accrue to enterprise software over consumer apps, as consumer utility has largely saturated while Ph.D.-level accuracy is critical for regulated corporate workflows. (25:00)
Contrary to expectations of platform disruption, Levie maintains that software distribution models remain unchanged: enterprise sales, PLG, and organic virality still dictate market winners regardless of agent sophistication. (30:00)
Detailed notes
## Background
Published by Aaron Levie on October 25, 2026, across RSS and web platforms as an episode of Speedrun Office Hours, this discussion examines how AI agents grant small startup teams unprecedented leverage to outrun corporate incumbents. According to the show description, the conversation explores how founders are transitioning from manual execution to orchestrating synthetic agents, why cheaper AI labor unlocks massive latent enterprise demand, and why traditional distribution playbooks remain essential despite rapid technological change.
## TL;DR
Box CEO Aaron Levie argues that AI agents represent the most radical structural shift in company building in decades, allowing teams of 5 to 10 people to match the output traditionally requiring hundreds of employees. Because manual code generation is commoditized, startup value has moved upstream toward spec-driven development, high-level strategy, and product taste. Levie anticipates hundreds of new billion-dollar companies emerging to automate specialized job functions across every vertical, with enterprise software capturing the bulk of future frontier model advances. However, he warns founders that product breakthroughs do not disrupt distribution: winning still requires disciplined, unglamorous go-to-market execution and owning target customer communities.
## Atmosphere & style
The conversation carries an energetic, pragmatic, and collegial tone, blending the enthusiasm of modern AI experimentation with grounded Silicon Valley operational wisdom. Levie brings self-deprecating humor about being an industry veteran since 2005 while maintaining genuine awe and curiosity toward the hyper-leveraged workflows of younger founders. The dynamic between host and guest feels like an insider peer-to-peer brainstorm, balancing bullish predictions about autonomous labor with sharp, realistic skepticism regarding enterprise friction and the myth of viral distribution.
## Bulletin
- `00:00` Box CEO Aaron Levie asserts that autonomous background AI agents provide founders with a 10x operational leverage multiplier, enabling 3- to 10-person startups to match the output traditionally requiring 50 to 200 employees.
- `05:00` Incumbent software giants face structural handicaps adopting agentic workflows due to legacy code debt, human-centric review pipelines, and organizational rigidity that induce multi-agent conflicts.
- `20:00` Levie models a 5,000-cell matrix mapping AI agents against specific economic job functions, arguing that dropping labor costs from hundreds of dollars to single digits will trigger a 50x expansion in demand for complex knowledge work.
- `25:00` Future foundation model gains will disproportionately accrue to enterprise software over consumer apps, as consumer utility has largely saturated while Ph.D.-level accuracy is critical for regulated corporate workflows.
- `30:00` Contrary to expectations of platform disruption, Levie maintains that software distribution models remain unchanged: enterprise sales, PLG, and organic virality still dictate market winners regardless of agent sophistication.
## Key takeaways
- `00:00` The structural mechanics of launching a startup have decoupled headcount from business output for the first time in 30 years.
- `05:00` Founders are shifting from manual keyboard-bound operators to managers, reviewers, and orchestrators of synthetic background labor.
- `05:00` Startups have a unique competitive window to outrun incumbents across both speed and scale simultaneously before larger firms can adapt.
- `10:00` As raw code generation commoditizes, value creation concentrates upstream around high-level strategy, product taste, and clear specification writing.
- `11:30` Mitigating AI context rot requires deliberate virtual specialization into modular subagents rather than building monolithic super-agents.
- `13:30` Autonomous coding tools represent the vanguard of a trillion-dollar restructuring across knowledge work that will bypass legacy tech vendors.
- `16:00` Background asynchronous agents provide 3x to 10x operational leverage, far surpassing the 10% to 20% efficiency gains of early autocomplete tools.
- `18:30` Because modern leverage amplifies individual capacity by an order of magnitude, co-founder quality and interpersonal friction matter more than ever.
- `20:00` A comprehensive matrix of 5,000 industry and functional roles offers a clear blueprint for standalone multi-billion-dollar agent businesses.
- `20:00` Compressed labor costs unlock massive latent enterprise demand previously excluded by high annual human employee salary floors.
- `20:00` The optimal founder strategy involves either personal operational mastery, pairing with a domain veteran, or shadowing end-users to deconstruct complex workflows.
- `25:00` Enterprise balance sheets will capture the bulk of future foundation model improvements because regulated corporate workflows demand Ph.D.-level precision.
- `25:00` Intentional multi-stage pivots from consumer software into enterprise software undermine company building; founders should target enterprise budgets directly.
- `30:00` Artificial intelligence does not reinvent distribution fundamentals; enterprise software still relies on disciplined enterprise sales and PLG motions.
- `31:38` Flashy product demos and social media virality cannot substitute for enduring audience ownership and dedicated go-to-market execution.
## Detailed walkthrough
### `00:00` The New Startup Playbook: 1995 vs. 2005 vs. 2025
The episode opens with the premise that AI agents are granting tiny teams unprecedented leverage, potentially minting hundreds of billion-dollar companies across every industry. Welcoming Box CEO Aaron Levie to *Speedrun Office Hours*, the host points out that while Levie is still relatively young, he has been running Box since 2005—meaning he has spent almost half of his life leading the company. Levie jokes that being reminded of that makes him feel ancient, even as he appreciates the host anchoring youth to his own age. Looking at today’s ecosystem, Levie expresses genuine awe at the influx of 20-year-old founders. When he and his co-founders dropped out during the middle of their junior year in 2005, dropping out to build a company in Silicon Valley was still a rare phenomenon—primarily spurred by Facebook—with only a few dozen peers doing the same. Today, dropping out has become the standard playbook.
Beyond founder demographics, Levie argues that the mechanics of building a company have undergone their most radical structural shift in decades. Comparing tech eras across 10-year increments, he explains that running a startup in 2005 felt almost identical to running one in 1995: both eras operated just before the cloud era took hold, requiring founders to physically go into data centers, rack servers, and write code morning and night. In stark contrast, building a company in 2025 compared to 2015 or 2005 is fundamentally different because of an AI-first orientation that decouples headcount from business output. While Levie admits that some modern founders seem “totally crazy and unhinged,” the operational leverage they possess is unprecedented. As Levie observes:
> “If I compare the 2025 startup to us in 2005, or let's say the 2015 startup, so if you just use 10-year increments, it is the most different that I've ever seen in the history of my adult life and doing technology now for 27 years, let's say.”
- Box was founded in 2005 after Levie and his co-founders dropped out in the middle of their junior year of college.
- In 1995 and 2005, startup execution required physical data center visits, racking physical hardware, and continuous manual coding.
- In 2025, an AI-first startup of 3, 5, or 10 people can achieve the output traditionally associated with 50, 100, or 200 employees.
`05:00` Founders as Agent Orchestrators and the Window Over Incumbents
Levie contrasts the day-to-day work of early Box with modern agentic workflows. In 2005, four co-founders lived and worked in a renovated garage, where their company’s entire rate of progress was strictly limited by typing speed—how fast they could manually punch out code, draft blog posts, create marketing materials, or manage servers in a single terminal. Today, founders and early employees are transitioning into managers of synthetic labor. Levie cites a startup founder who spends his days reviewing thousands of lines of code generated by background agents. Because an engineer is no longer constrained to typing into one IDE at a time, their role shifts from primary author to editor, reviewer, and orchestrator—yielding roughly a 10x multiplier in leverage.
While Levie cautions that agents are not a silver bullet—noting that many founders still compare interacting with agents to playing a “slot machine” that requires repeated rerolls to get clean results—he emphasizes that the competitive gap between nimble startups and resource-rich incumbents is shrinking fast. Startups have historically compensated for lack of resources with pure speed, but incumbents could always deploy more capital and bodies. Now, as small teams use agents to scale engineering capacity tenfold, they can outrun incumbents across scale and speed simultaneously. Incumbents struggle to counter this because their complex internal processes and existing workflows are too rigid to adapt quickly. While this pattern began in software engineering, Levie expects it to spread rapidly to marketing, legal work, and outbound sales. The host notes that as execution becomes cheap, the pressure shifts to high-level strategy and taste: when execution barriers vanish, the decisions made on the strategic chessboard become make-or-break. As Levie highlights regarding the incumbent vulnerability:
> “So you have a window where startups can actually outrun these big companies in a way that was just never economically feasible before.”
- Early Box operations were strictly rate-limited by the physical typing speed of the four founders working in a converted garage.
- Editing and orchestrating AI-generated code provides roughly 10x the operational leverage of writing code manually from scratch.
- The primary developer bottleneck has shifted from raw code production to code review, prompt orchestration, and architectural integration.
- Incumbents face an organizational handicap: entrenched workflows and legacy processes prevent them from retooling around autonomous agents as rapidly as greenfield startups.
`10:00` Spec-Driven Development and the Premium on Taste
The shift toward AI-assisted engineering has inverted the traditional startup skill stack. Software creation now hinges on crafting precise prompts and in-depth specifications that allow agents and subagents to navigate and build within distinct parts of a codebase. Aaron Levie observes that this rise of "spec-driven development" has quietly supercharged product managers: the stereotypical "pointy-haired Dilbert product manager" now possesses the direct leverage to produce real, functional software. When writing an articulate specification becomes the primary bottleneck to shipping software, the ability to clearly define what needs to be built confers extraordinary power.
Because the mechanical act of hacking an MVP together is becoming commoditized, speed of raw execution no longer guarantees competitive differentiation. Instead, value creation has shifted upstream to strategic rigor, product sensibility, and aesthetics. Founders must grapple with fundamental questions: whether an authentic market opportunity exists, whether the product looks exceptional, and whether the team possesses genuine taste. As Aaron Levie puts it: “Well now actually because the hacking something together is fairly commoditized, the premium is have you thought through what you actually trying to build is a real market opportunity for it is it a good idea Does it look good Does it have taste Those are now the new forms of value creation.”
```mermaid
flowchart LR
accTitle: The Spec-Driven Engineering Loop
accDescr: Diagram illustrating how high-level specifications and taste guide subagents to produce code without traditional manual hacking.
subgraph Ideation ["Strategic Formulation"]
A["Product Spec & Market Strategy"] --> B["Design & Taste Verification"]
end
subgraph Execution ["Agentic Execution"]
B --> C["Specialized Subagents"]
C --> D["Isolated Codebase Modules"]
end
```
`11:30` Combating Context Rot Through Specialized Subagents
When choosing how to structure agent workflows, startups face the problem of "AI context rot"—the degradation of model performance when an agent is overloaded with an entire codebase and disparate objectives. Rather than relying on a single, monolithic "super agent" or "ultimate fighter" that attempts to handle every task, effective teams enforce functional boundaries. Levie points to platforms like Claude Code, which enable virtual specialization by letting developers spin up subagents dedicated to discrete slices of a codebase or specific stages of an operational workflow.
Levie clarifies that addressing context rot is not about accumulating dozens of separate point-solution SaaS tools, but rather about deliberately architecting specialization. Because the all-knowing super-agent is still a myth, founders must segment tasks so language models can operate within clear context boundaries. Best practices remain in flux; interviewing five different startups yields five distinct approaches, ranging from modular subagents to tightly scoped prompt chains. For startups, this creates a major opening: by engineering specialized tools and tailored workflows that incumbents cannot easily adopt, new companies can capture critical ground across the development stack.
- Context rot is solved not by tool bloat, but through virtual or tangible agent specialization.
- Tools like Claude Code demonstrate this model by assigning subagents to specific modules or functional workflows.
- “It’s that we have a very real issue, which is this, you know, like super powerful, super agent doesn't exist. It doesn't work. And so you do have to have some degree of specialization and that specialization can come, you know, in a variety of packages.” — Aaron Levie
`13:30` Coding Agents and the Trillion-Dollar Knowledge Work Market
The productivity explosion seen in AI-native startups points toward a total restructuring of the broader software landscape. Categories of knowledge work that historically lacked bespoke software are now being automated through autonomous agents. The first domain to achieve critical mass is software engineering itself, where an influx of startups—including Replit, Cursor, Windsurf, Lovable, and Cognition—has rewritten the rules of the developer ecosystem. A decade ago, GitHub stood as virtually the only massive venture exit in developer tools, making the idea of tens of billions of dollars in coding tool market cap seem preposterous.
Levie argues that the current valuation surge is merely scratching the surface. He contends that AI coding tools will generate hundreds of billions of dollars in enterprise value that will bypass legacy tech incumbents because the total addressable market is expanding so rapidly. He emphasizes that the dynamic observed in coding will replicate across every high-skill vertical: “Do that for life sciences, do that for healthcare, do that for legal work, do that for consulting. There is going to be new AI agent companies in every single one of those spaces. And not only will they outrun the traditional software incumbent companies, they're going to outrun the traditional services companies in those spaces as well.”
| Dimension | Legacy Software Era | AI Agent Era |
|:--- |:--- |:--- |
| **Primary Value Engine** | Manual software engineering & seat licenses | Autonomous agents performing domain knowledge work |
| **Developer Tools Exit Precedents** | Sparse (predominantly GitHub acquisition) | Multi-billion market caps across Replit, Cursor, Windsurf, Cognition |
| **Competitors Disrupted** | Traditional enterprise software vendors | Software incumbents *and* traditional professional service firms |
| **Target Verticals** | Standardized corporate business processes | Life sciences, healthcare diagnostics, legal services, management consulting |
`16:00` 10x Background Leverage vs. Incumbent Drag
When Levie speaks with early-stage founders today, he finds himself learning from them roughly 70% of the time, forced to constantly recalibrate his assumptions about operational leverage. Just a year prior, most conversations revolved around 10% to 20% efficiency gains from GitHub Copilot's autocomplete or inline question answering. Today’s top founders operate with 3x to 10x leverage by delegating entire tasks to autonomous background agents that run asynchronously across the codebase.
Large incumbents struggle to adopt this operational cadence because their organizational machinery resists it. Enterprise codebases frequently lack the pristine documentation agents need to operate autonomously. Moreover, their engineering methodologies, review pipelines, and collaboration frameworks are designed to coordinate humans, not to handle dozens of background agents refactoring code simultaneously without generating massive merge conflicts. Small, greenfield startups possess an innate structural advantage: free from legacy code debt and coordination overhead, a two- or three-person engineering team can operate with the output capacity of an organization hundreds of times its size. Levie’s focus within Box is determining how to bridge these cutting-edge startup patterns into an established enterprise environment.
- Autocomplete and typeahead tools delivered modest 10–20% productivity bumps; asynchronous background agents yield 3x–10x improvements.
- Incumbent engineering teams face structural roadblocks: poorly documented codebases, incompatible tool stacks, and human-centric revision processes prone to multi-agent conflicts.
- “It actually is something that benefits a small team where they're not going to have all these conflicts happening at all time across all these agents. And so this is something that naturally is advantaged for the smaller startup to be able to move quickly on.” — Aaron Levie
`18:30` Timeless Company Building and Choosing Co-Founders
Reflecting on what advice he would give to a young founder dropping out of college today, Levie underscores that while the leverage formula has evolved, the foundational rules of entrepreneurship remain durable across technology cycles. Because autonomous agents multiply the capacity of individuals to an unprecedented degree, co-founder selection has become vastly more consequential. A tiny founding team equipped with modern agent architectures commands the equivalent leverage of hundreds or thousands of workers, meaning any weakness, friction, or lack of skill in the core team is magnified proportionally.
While some entrepreneurs argue for pursuing obscure, heavily regulated, or boring niche markets to achieve regulatory capture early on, Levie maintains that founders must first decouple upstream ideation from downstream execution. The principles of picking exceptional co-founders, maintaining strategic clarity, and cultivating taste are timeless requirements that precede product selection or regulatory maneuvering. As Levie notes regarding the magnified importance of founding personnel: “To your exact point, team probably matters more than ever before because of the leverage on your co-founders or co-founding team.”
`20:00` The 5,000-Cell Agent Matrix and Elastic Demand for AI Labor
Aaron Levie envisions the next wave of tech giants emerging from a straightforward formula: identifying a specific job function in the economy, building a dedicated AI agent to perform it, and selling that agent to enterprises, service firms, and systems integrators. He suggests constructing a massive matrix across every industry and job function—a grid of roughly 5,000 cells mapping roles like life sciences sales representatives or financial services compliance attorneys. While some cells may be constrained by total addressable market (TAM), others will support standalone businesses valued at $1 billion, $5 billion, or $10 billion and beyond. For founders deciding what to build, Levie highlights advice from Y Combinator partner Jared Friedman: “Go do the job of what you want the agent to go and automate and be like the expert in how to do that job.” The ideal founder is someone who has already performed the work personally; the next best approach is finding a co-founder with that operational background, or directly shadowing practitioners full-time to master the nuances of workflows like FDA regulatory reviews or legal case analyses.
Beyond automating existing spend, Levie argues that analysts consistently misjudge the market size for AI agents because they overlook supply-demand elasticity. In many economic sectors, high costs artificially depress demand. High-end legal review, for instance, commands $1,000 to $2,000 per hour, forcing clients to ration legal services. If an AI agent can execute that same work at $5 or $10 an hour, the market will not simply contract; it will experience an explosion in volume as ordinary individuals and small firms file patents or conduct exhaustive contract reviews that were previously cost-prohibitive. As Levie summarizes the dynamic: “It will turn out that there's more categories than we think, where right now the supply-demand equilibrium in the economy is actually not sort of as perfect as we would have imagined, where 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.”
Inside Box, Levie observes identical dynamics during internal budget planning. Numerous high-value operational projects remain permanently stranded at the bottom of corporate stack-ranking lists simply because hiring a full-time employee costs $100,000 to $300,000. An AI agent priced at $1,000 removes the traditional minimum floor price, unlocking return on investment with virtually no ceiling. Once companies deploy agents against these overlooked bottlenecks, unexpected efficiencies emerge, prompting leaders to allocate even more capital to those functional areas. This dynamic prompts a running joke about waves of eager founders descending on back-office operations:
- **The Founder Playbook Hierarchy**:
1. Automate a job you have personally performed for years.
2. Team up with an operational domain expert as a co-founder.
3. Spend weeks shadowing professionals to deconstruct workflows line by line.
- **Budget Stack-Ranking Dynamics**: Traditional corporate headcount carries a steep entry barrier ($100,000–$300,000+ per full-time equivalent), leaving lower-priority projects unstaffed; software agents priced around $1,000 eliminate the minimum cost threshold entirely.
- **Back-Office Infiltration Joke**: The hosts anticipate an unexpected influx of ambitious twenty-somethings suddenly taking administrative desk jobs in FDA-regulated biopharmas, greeting confused managers with: *"Oh, Aaron sent me."*
```mermaid
flowchart TD
accTitle: Labor Cost Compression Unlocking Latent Enterprise Demand
accDescr: A flowchart showing how reducing labor costs by an order of magnitude transforms backlogged corporate tasks into active AI agent deployments.
subgraph Traditional["Traditional Human Labor"]
A["High Cost Floor<br/>($100k-$300k/year)"] --> B["Strict Budget Prioritization"]
B --> C["Latent Enterprise Tasks<br/>(Cut at Bottom of Stack Rank)"]
end
subgraph Agentic["Agentic Labor Model"]
D["Low Marginal Cost<br/>($1k Agent Deployment)"] --> E["Eliminates Minimum Floor"]
E --> F["50x Expansion in Volume<br/>(Patent filings, deep reviews)"]
end
C -.->|"Compressed 10x in Cost"| D
F --> G["Bottlenecks Cleared &<br/>ROI Ceiling Removed"]
```
---
`25:00` Why Model Progress Amortizes into the Enterprise
Reflecting on the gap between technical builders and subject matter experts, Levie proposes a matchmaking mechanism to pair "central casting" technical founders—such as MIT dropouts who know how to architect agent frameworks—with domain veterans plucked directly from companies like Pfizer who understand intricate regulatory approval pipelines. When asked whether founders should cut their teeth building consumer products before pivoting into enterprise software (a historical pattern echoed by Box's own evolution and past advice from venture capitalists like Marc Andreessen), Levie pushes back on intentional pivoting: “I don't think I would plan a pivot. That kind of breaks the definition of a pivot.” While consumer AI will continue to produce interesting applications, Levie contends that enterprise software will capture the lion's share of total addressable market value simply because artificial intelligence is fundamentally a productivity engine.
To illustrate why future foundation model advancements will disproportionately benefit enterprise software rather than consumer applications, Levie cites an observation made by Anthropic CEO Dario Amodei. If a frontier model advances from the capability of an undergraduate chemistry student to that of a chemistry Ph.D., an average consumer using Claude or ChatGPT will scarcely notice the difference. A parent generating bedtime stories, generating family songs, or asking basic everyday questions has already reached full capability saturation with existing frontier models like GPT-4. However, that exact same leap to Ph.D.-level precision delivers colossal financial value to a pharmaceutical giant like Pfizer screening drug candidates, or to a law firm auditing an intricate contract where current models still leave dozens of critical errors.
Levie concludes that model progress will inevitably amortize into enterprise software balance sheets: “By definition, most AI progress is going to now amortize into enterprise use cases.” While novel consumer companies will still be founded, their breakthroughs will stem from clever packaging and user experiences rather than raw model intelligence upgrades. By contrast, deep enterprise workflows require absolute factual accuracy, making enterprise software the primary economic beneficiary as frontier models approach graduate-level domain mastery—a realization that naturally shifts the founder's challenge from raw product intelligence toward the perennial hurdle of enterprise distribution.
- **The Co-Founder Matchmaking Model**: Pairing MIT-pedigree agent engineers with industry veterans (e.g., Pfizer regulatory specialists) to spin up dozens of category-specific vertical agent startups.
- **The Ph.D. Capability Threshold**: Consumer chat queries already find existing frontier models "good enough" for everyday creative or informational needs; enterprise use cases (e.g., legal review finding 70 contract flaws) urgently require graduate-level analytical rigor.
- **The Limits of Consumer Lift**: Aaron notes that for his personal life and parenting—making songs, illustrations, and bedtime stories for his six-year-old—AI capabilities have already hit functional saturation.
- **The Pivot Fallacy**: Starting in consumer software with the premeditated plan of pivoting to enterprise defeats the nature of a pivot; startups should aim directly at large enterprise productivity budgets from day one.
`30:00` Why AI Will Not Disrupt Tech Distribution
The discussion turns to go-to-market dynamics and whether artificial intelligence will reinvent distribution the way it has reshaped product capabilities. The host observes that while AI has yielded obvious product insights, distribution models appear largely unchanged, asking whether the ecosystem is simply waiting for underlying platforms to mature or if the transition will be as radical as moving from radio to television. Aaron Levie offers an intentionally grounded, contrarian take on the matter.
Levie admits that his prediction might sound unexciting or even prove wrong down the line, but he maintains that the fundamentals of software distribution will remain intact. While novel discovery channels will surface—such as landing featured placement inside an OpenAI app store or agent marketplace—the overarching playbook remains unchanged. Consumer apps still live or die by organic virality, enterprise software still demands a disciplined enterprise sales machine, and software targeting technically savvy, self-serve buyers still relies on product-led growth (PLG) paired with an enterprise sales layer. As Levie summarizes:
> “I don't think this is a fundamental transformation or disruption of the distribution model of tech.”
```mermaid
flowchart TD
accTitle: Tech Distribution Models in the AI Era
accDescr: Aaron Levie's categorization of distribution strategies across market types.
subgraph Consumer ["Consumer Market"]
C1["Consumer AI Apps"] --> C2["Virality & Organic Reach"]
end
subgraph Enterprise ["Enterprise Market"]
E1["B2B / Industry Agents"] --> E2["Dedicated Go-to-Market Engine"]
end
```
`31:38` "If You Build It, They Will Not Come": Owning the Audience
Expanding on the distribution dilemma, Levie warns founders that building an extraordinary agent is nowhere near enough to guarantee business survival. A brilliant technical implementation is easily drowned out by competing priorities and thousands of rival startups pitching similar tools. Levie cautions against mistaking early social media hype or flashy demo videos for genuine commercial traction, emphasizing that real distribution is an unglamorous, continuous grind. As Levie warns:
> “If you think that because you built this amazing agent, that now you've solved your problem, like that is like totally you're going to fail.”
True distribution, in Levie’s view, requires actively capturing and maintaining an audience that is deeply passionate about the specific problem being solved. Startups must establish direct access to target users across community channels like LinkedIn, X, and Reddit, executing a sustained "bear hug" around their target industry through dedicated marketing, sales outreach, and PR. The host humorously predicts that startup founders will soon mandate industry domain experience for every hire—mirroring the classic belief that everyone should work in customer service to understand customer needs. Levie agrees that founders who disregard these traditional distribution fundamentals leave themselves completely exposed to rivals who simply execute better on reaching customers, closing out the conversation on the necessity of disciplined commercial execution.
- **Demo virality vs. real distribution:** A viral video that pulls in views and social likes is merely a one-time marketing stunt, not a sustainable distribution pipeline.
- **Audience ownership:** Founders must build direct, enduring relationships with specialized user communities where buyers actually congregate.
- **Vulnerability to competitors:** Companies with superior product capabilities will still be crushed by competitors possessing stronger distribution engines and closer customer proximity.
## Notable quotes
> `25:00` By definition, most AI progress is going to now amortize into enterprise use cases. — Aaron Levie
> `25:00` I don't think I would plan a pivot. That kind of breaks the definition of a pivot. — Aaron Levie
## Mentioned
Aaron Levie, Box, Speedrun Office Hours, Claude Code, Replit, Cursor, Windsurf, Lovable, Cognition, GitHub, GitHub Copilot, Y Combinator, Jared Friedman, Pfizer, Marc Andreessen, Anthropic, Dario Amodei, Claude, ChatGPT, GPT-4, OpenAI, LinkedIn, X, Reddit
## Video description
Can a 10-person startup outperform a company with hundreds of employees? Box CEO Aaron Levie argues AI agents are giving small teams a level of leverage that simply wasn't economically possible before, creating a window for startups to outrun incumbents.
In this episode of Speedrun Office Hours, Aaron explains why founders are shifting from doing the work themselves to orchestrating agents, and why that puts a new premium on strategy, taste, and knowing exactly what to build. He makes the case that nearly every industry and job function could support specialized AI agents, potentially creating hundreds of new billion-dollar companies.
They also discuss why enterprise AI may capture much of the next wave of model improvements, how dramatically cheaper AI labor could unlock demand that doesn't exist today, and one thing Aaron thinks AI won't change: distribution. If you build it, they still won't necessarily come.
Box CEO Aaron Levie argues that autonomous AI agents have decoupled headcount from business output, enabling tiny engineering teams to outpace large enterprise incumbents. However, Levie contends that as software development becomes commoditized, long-term competitive advantage shifts upstream toward strategic taste, domain specialization, and traditional distribution execution.
Headcount Decoupling and the Incumbent Asymmetry (0:00)
Levie contrasts the 1995 and 2005 startup environments—where progress was rate-limited by manual server racking and physical typing speed—with 2025's AI-first model. Early employees and founders now act as editors and orchestrators managing asynchronous background agents across codebases. This approach yields an estimated 3x to 10x multiplier in operational leverage, far exceeding the 10% to 20% efficiency gains provided by earlier autocomplete tools.
This shift creates an operational window for greenfield startups over established incumbents. Small teams of three to ten people can match output previously requiring 50 to 200 employees. Large enterprises struggle to deploy agents at this scale because they are constrained by legacy codebases, poor documentation, and collaboration processes designed for humans that frequently encounter merge conflicts under multi-agent code generation.
Spec-Driven Engineering and Context Specialization (10:00)
As the mechanical task of writing code becomes commoditized, Levie observes an industry pivot toward spec-driven development. Crafting precise, detailed functional specifications has become the core engineering bottleneck, granting product managers the leverage to ship functional code directly. With basic execution barriers lowered, Levie argues that value creation increasingly hinges on taste, product aesthetics, and strategic clarity.
To execute effectively, teams must also address AI context rot—the performance degradation that occurs when a single model is overloaded with an entire codebase and conflicting goals. Because monolithic super-agents do not yet work reliably, Levie emphasizes that startups must enforce virtual or structural specialization, using tools like Claude Code to assign subagents to isolated functional modules.
Labor Price Deflation and Latent Demand (20:00)
Levie outlines a major startup opportunity in mapping approximately 5,000 distinct industry job functions and building dedicated autonomous agents to automate them. The most successful founders, he suggests, either automate jobs they have personally performed or partner directly with domain veterans who understand complex, highly regulated workflows like FDA compliance reviews.
Furthermore, Levie argues that market analysts consistently underestimate agent total addressable market by overlooking price elasticity. Traditional corporate projects are frequently abandoned because human full-time equivalents cost $100,000 to $300,000 annually. Deploying an agent at approximately $1,000 removes the traditional budget floor, unlocking substantial latent demand in areas like legal document review where high hourly billing rates previously forced clients to ration services.
Enterprise Amortization of Frontier Model Progress (25:00)
Levie contends that future advancements in foundation models will accrue almost entirely to enterprise software rather than consumer applications. Citing Anthropic CEO Dario Amodei, Levie notes that advancing a model from an undergraduate to a Ph.D. level in domain reasoning goes largely unnoticed by consumers generating casual text or images, as existing models already satisfy everyday consumer use cases.
By contrast, mission-critical enterprise workflows—such as pharmaceutical candidate screening or legal contract audits—require absolute accuracy and cannot tolerate hallucinations. Because deep enterprise processes depend on graduate-level analytical precision, Levie asserts that model intelligence gains will amortize directly into B2B software, making deliberate pivots from consumer to enterprise an inefficient startup strategy.
The Unchanged Fundamentals of Tech Distribution (30:00)
Despite artificial intelligence transforming software creation, Levie maintains that go-to-market dynamics remain fundamentally unchanged. Consumer applications still require organic virality, self-serve technical tools still depend on product-led growth paired with enterprise sales, and B2B products still demand disciplined direct sales teams.
Levie warns founders against confusing viral social media demos with genuine commercial traction. Even an exceptional agentic product will fail without sustained audience ownership and direct engagement in target professional communities, as competitors with superior traditional distribution engines will reliably capture the market.
Levie concludes that while autonomous agents compress the cost and labor required to build software, they heighten the stakes of foundational business decisions. With technical execution commoditized, durable startup value concentrates in co-founder selection, deep domain accuracy, and disciplined enterprise go-to-market execution.
## Background
本期内容由 Box 首席执行官 Aaron Levie 于 2026 年 10 月 25 日在 RSS / 网页平台的《Speedrun Office Hours》节目中发布。页面简介指出，对谈聚焦于 10 人初创团队如何借助 AI Agent 获得前所未有的生产力杠杆并超越大企业，探讨了创始人从亲力亲为转向智能体调度的角色转变，以及企业级 AI、垂直领域机遇和始终未变的分发（Distribution）本质。
## TL;DR
Box 创始人兼 CEO Aaron Levie 深入剖析了 AI 原生时代创业范式的底层重构，指出 3 至 10 人的微型团队如今能释放出以往数百人的工程与业务产能。创始人的核心职责已从纯手写代码转向调度与审阅 Agent，战略眼光、市场洞察与产品品味成为全新的核心溢价。他预测各行各业的 5,000 个细分职能将被垂直 AI 智能体全面重塑，劳动成本的骤降将激发数十倍的潜在需求。同时他清醒地警告，前沿模型智力的跃升红利绝大多数将被企业端吸收，且 AI 丝毫没有颠覆传统的分发逻辑，“做出来用户就会来”是致命幻觉，团队依然必须扎实构建 GTM 与核心受众壁垒。
## Atmosphere & style
整场对话洋溢着硅谷老兵对新一代技术浪潮的由衷兴奋与谦逊探索，Levie 既坦言自己有大量时间是在向年轻创始人请教，又以二十年创业积淀展现出穿透技术狂热的极其老练的商业判断。访谈节奏明快、谈笑风生，在拆解代码规范与 Agent 架构等硬核工程细节之余，穿插着幽默的自嘲与生动的行业隐喻。整体听感兼具高维度的战略推演与冷峻的现实警醒，在极度乐观的技术展望与严谨务实的商业铁律之间达成了绝妙平衡。
## Bulletin
- `00:00` Box CEO Aaron Levie 指出，AI 原生架构带来了近 30 年来最剧烈的创业范式变革，3 到 10 人的微型团队现已能释放出传统 50 至 200 人规模的产能。
- `10:00` 随着 AI 降低代码编写门槛，软件开发的竞争溢价已从单纯的“黑客拼凑速度”彻底转向战略洞察、真实市场机会判断以及产品品味。
- `20:00` 通过全经济体矩阵交叉分析，Levie 预言未来几乎每个行业的具体岗位都能诞生专用 AI Agent，整体市场有望催生数千亿美元的新增市值。
- `25:00` 大模型智力从本科跃升至博士水平的边际收益主要被容错率极低的企业级场景所吸收，日常消费端的需求目前已基本达到饱和。
- `30:00` AI 并未颠覆商业软件的分发逻辑，仅凭技术构建产品而缺乏 GTM 与深度垂类受众壁垒的初创公司必然面临淘汰。
## Key takeaways
- `05:00` AI Agent 正在迅速抹平初创团队与大厂之间的资源鸿沟，为敏捷的小型团队创造了前所未有的超车时间窗口。
- `05:00` 创始人工作角色正从亲自敲击代码的“一线编写者”转变为在后台并行调度、整合与审阅多个 AI Agent 的“管理者”。
- `10:00` 软件工程迈入“规范驱动开发”（spec-driven development）时代，详尽的设计规范与清晰的提示词编排成为交付的核心。
- `11:42` 全能型超级 Agent 在工程现实中并不成立，必须依靠垂直解耦的子 Agent 矩阵来规避“AI 上下文腐烂”（AI context rot）问题。
- `13:40` AI 编程赛道只是垂直知识工作被智能体改造的冰山一角，类似模式正向生命科学、法律、咨询等高价值领域快速平移。
- `16:03` 成熟大企业受制于代码库冲突、组织阻力与既有开发规范，无法像小团队那样无包袱地全面释放背景自主 Agent 的杠杆倍数。
- `18:24` 在 AI 极大放大个人产出的时代，初创团队联合创始人的品味、视野与默契不仅没有被技术稀释，反而变得前所未有的关键。
- `20:00` 打造高壁垒垂直 Agent 的最佳路径是创始人具备一线从业经验，或引入垂类专家合伙人，对真实业务流进行逆向拆解。
- `22:20` 劳动力成本暴跌 90% 将释放巨大的弹性需求，以往因人力成本过高而被企业预算搁置的业务需求将被 Agent 彻底激活。
- `25:00` 前沿模型能力的持续升级对日常消费级场景意义有限，但能直接解决企业级复杂合同审查与研发等容错率极低的痛点。
- `30:00` AI 没有颠覆商业分发的底层逻辑，To C 依然靠裂变传播，To B 依然依赖扎实的销售机器与 GTM 引擎。
- `31:40` “做出来用户自然会来”是致命幻觉，爆款 Demo 视频不等于真实分发，初创公司必须建立起持续触达核心垂类受众的护城河。
## Detailed walkthrough
### `00:00` 创业范式的世纪巨变：从车库手敲代码到 10 人敌百人
在访谈开场，主持人调侃 Box 首席执行官 Aaron Levie 虽然实际年龄依然年轻，但自 2005 年创立 Box 至今已有整整二十年，几乎占据了他人生的一半时间。当被问及如今的创业环境与当年有何不同时，Levie 笑称自己最大的感触其实是“觉得自己老了”。他回忆道，2005 年他和合伙人在大三中途退学创办 Box 时，二十岁辍学创业在硅谷还是极少数人的异类举动，当时整个硅谷可能只有几十个人这么做，虽然 Facebook 的成功让辍学开始受到关注，但这在当时仍属极其罕见的个例。而放眼当下，二十岁退学创业几乎已经变成了年轻创业者的标配打法，这种新生代创业者的蓬勃活力令人惊叹。
```mermaid
flowchart TD
accTitle: 创业生产力范式演变
accDescr: 对比1995至2015年传统创业模式与2025年AI原生创业模式的生产力差异。
subgraph TRAD["传统模式 (1995 - 2015)"]
direction TB
A1["物理限制: 单一键盘/IDE"] --> B1["人力规模线性产出: 3-4人团队"]
B1 --> C1["大公司依靠100倍人力资源形成碾压"]
end
subgraph AI["AI 原生模式 (2025+)"]
direction TB
A2["多后台并发: 部署多个 AI Agent"] --> B2["杠杆倍增: 3-10人实现50-200人产出"]
B2 --> C2["大公司受制于固定流程，小团队实现敏捷超越"]
end
```
谈到技术变革的周期，Levie 指出，如果以十年为一个跨度来审视创业公司的运作方式，当下的变化是近三十年来最为剧烈的一次。在 2005 年创办 Box 时，云计算浪潮尚未完全爆发，创始团队需要亲自跑去数据中心上架服务器，从早到晚手动敲击代码。如果把 1995 年的初创公司与 2005 年相比，两者的运作模式几乎完全一致，即便对比 2015 年，核心逻辑也没有发生颠覆。然而，2025 年的初创公司与以往任何时代都彻底不同。Aaron Levie 总结道：“If I compare the 2025 startup to us in 2005, or let's say the 2015 startup, it is the most different that I've ever seen in the history of my adult life and doing technology now for 27 years, let's say. And it's simply because of this AI-first orientation, where you can be a three, five, 10-person team and have the output of a 50 or 100 or 200-person team.” 这种全新的“AI 优先”定位，彻底重构了企业的运行逻辑，也孕育着诞生数百家估值在 10 亿、50 亿乃至 100 亿美元以上新兴巨头的无限可能。
- 2005 年硅谷辍学创业者仅有几十人，而如今已成为年轻创业者的常见选择。
- 1995 年与 2005 年的创业模式几乎完全相同，依赖物理机房服务器与纯手工代码构建。
- 2025 年的初创公司凭借 AI 原生架构，能够让 3 到 10 人的微型团队释放出以往 50 至 200 人规模的产能。
---
### `05:00` 创始人角色剧变：从键盘打工人到 AI Agent 的调度者与审阅者
深入剖析当下初创团队的日常运作，Levie 描绘了创始人工作方式的根本性断层。在创立 Box 之初，整个团队的推进速度百分之百受限于手指敲击键盘的速度，无论是编写代码、撰写营销资产、发布博客文章还是维护服务器，四个合伙人蜗居在由车库改建的紧凑空间内，执行效率被单人操作终端的物理极限死死卡住。而如今与新兴创始人交流时，Levie 发现了一种普遍趋势：创始人和早期员工的角色正在转向“工作的管理者”（Managers of work），具体执行的大量工作则交由 AI Agent 完成。人类创业者不再被编码和创作的速度所限，而是把精力集中在编排工作流与下达任务上。
Levie 举例提到，最近一位创业团队的创始人向他透露，自己现在每天主要的工作就是审阅 Agent 自动生成的数千行代码。Levie 指出，担任“代码编辑器与审阅者”所获得的生产力杠杆，是传统一线“编写者”的十倍以上。程序员受生理限制一次只能在一个 IDE 中敲打一段代码，但作为管理者，创始人可以在后台同时调度多个 Agent 并行作业，待它们全部完工后再统一审阅合并。此时，初创公司的研发瓶颈不再是“人类能写多少代码”，而是“人类审阅、整合并触发新 Agent 的速度”。尽管目前的 Agent 并不完美，很多人将其比作“拉老虎机”（Slot machine）——必须反复调试提示词、等待生成并看运气，但只要 Agent 生成的代码质量达到标准并符合代码规范，创始人只需稍微修剪调整，就意味着初创公司瞬间获得了十倍的工程产能。
这种生产力跃迁正在从根本上改写初创公司与大厂之间的竞争法则。在以往的商业竞争中，两三人的初创公司虽然动作敏捷，但面对拥有百倍人力的行业巨头时，巨头始终能通过调集海量人力资源形成压制；而现在，Agent 正在迅速抹平这种资源差距。谈到大企业的应对困境与初创团队的机会窗口，Aaron Levie 直言：“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.” 大公司受制于根深蒂固的既有业务流程与组织阻力，转型极其缓慢，这给小型团队留出了前所未有的超车时间窗口。这种由代码领域开启的 Agent 化革命，正快速蔓延至市场营销、法律合规与出海销售等各个环节。随着基础执行工作被大幅压缩，商业竞争的重心正在全面倒逼创始人走向纯粹的策略与创意层面，在低执行成本下，棋盘上的战略决策与审美趣味正变得前所未有地关键。
- 创作者转向审阅者：创始人从亲自编写代码转变为在后台并发调度多个 Agent，工作杠杆率提升 10 倍。
- 竞争壁垒中和：初创团队攻打大型对手时原本天然存在的“资源规模劣势”，正在被 AI Agent 迅速填平。
- 巨头组织阻力：大型既得利益者由于既有工作流过于繁冗固化，无法快速切换到全 Agent 模式，为敏捷小团队创造了宝贵的超车窗口。
- 领域扩展路径：从早期的软件工程代码生成，逐步外溢渗透至市场宣发、法务合同处理以及 Outbound 销售获客等多个业务职能。
### `10:00` 规范驱动开发与 AI 时代的真正溢价：策略与品味
Aaron Levie 深入剖析了当今初创企业构建软件的全新范式——“规范驱动开发”（spec-driven development）。在这种模式下，软件工程的重心转移到撰写极度清晰的提示词（prompts）或详尽的设计规范上，并让 AI Agent 深入理解代码库，在系统各处调动专门的子 Agent 协同作业。Aaron Levie 幽默地借用漫画《呆伯特》（Dilbert）里的桥段调侃道，那个曾经被视为外行的“尖头产品经理”，如今终于获得了亲自动手造物的超能力：“只要你熬得够久，他们就会变得极其强大。”
然而，当技术实现的门槛被抹平时，创业的核心壁垒也发生了彻底逆转。过去，初创企业的溢价在于“你能多快拼凑出一个可用版本”（hack something together）；而如今，代码拼凑已经高度商品化，真正的稀缺能力变成了战略眼光、市场洞察与品味——你是否深思熟虑过自己究竟要造什么、这是否是一个真正的市场机会、产品是否美观且富有品味。正如 Aaron Levie 明确断言：“现在，拼凑出东西基本上已经商品化了，溢价在于你是否深思熟虑过自己到底想造什么、它是不是一个真正的市场机会、它是不是一个好主意、它看起来美不美、它有没有品味。这些现在成了全新的价值创造形式。”那些对这些核心要素保持极高追求的公司，才会在未来的竞争中脱颖而出。
```mermaid
flowchart LR
accTitle: 软件开发范式与价值转移
accDescr: 对比传统黑客拼凑模式与现代规范驱动开发模式的核心差异及价值重心转变
subgraph OLD["过去：拼凑驱动模式"]
A["快速黑客拼凑<br/>(Hack Together)"] --> B["早期可用产品"]
B --> C["高溢价：编码速度与工程实现"]
end
subgraph NEW["现在：规范驱动模式"]
D["清晰规范与提示词<br/>(Spec-Driven)"] --> E["编排子 Agent 矩阵"]
E --> F["高溢价：战略、市场机会与品味"]
end
```
- 规范驱动开发赋予产品经理前所未有的工程交付能力，缩短了从想法到实现的原型链条。
- 编程和粗糙实现已被 Agent 商品化，单纯比拼敲代码速度不再具备竞争壁垒。
- 新商业价值的决定性因子转向：产品战略、市场机会清晰度、审美设计与产品品味。
### `11:42` 破解“AI 上下文腐烂”：虚拟专业化与子 Agent 架构
针对初创团队如何搭建 Agent 架构、避免掉入架构泥潭的提问，对话转向了“AI 上下文腐烂”（AI context rot）的治理困境。主持人提到，行业中一种解决思路是放弃寻找无所不能的“全能终极战士”（ultimate fighter），转而构建一支分工明确的“复仇者联盟”式的 Agent 团队。Aaron Levie 指出，核心并不在于创业公司用了多少种外部工具，而在于必须以某种形式实现“专业化”（specialization）。
以 Claude Code 为例，它通过允许开发者针对代码库的特定模块或特定工作流创建专门的子 Agent，从而在虚拟层面实现了专业化分工，有效化解了上下文腐烂问题。Aaron Levie 直言：“因为我们面临一个非常现实的问题，那就是：根本不存在什么超级强大、全能的超级 Agent。那行不通。”因此，专业化是必然趋势，但具体实现路径百花齐放——问三到五家初创公司就会得到三到五种不同的答案，目前整个行业都在摸索最佳实践。
- “全能超级 Agent”（super-agent）在工程现实中并不成立，必须依赖子 Agent 进行业务与代码模块解耦。
- Claude Code 展现了通过子 Agent 隔离代码上下文以防止“上下文腐烂”的标准范式。
- 当前 Agent 架构尚无通用定式，行业仍处于最佳工程实践的密集探索期。
### `13:40` 垂直领域知识工作重塑：数千亿美元的新市场机遇
在探讨了 AI 如何让初创企业跑出比巨头快 10 倍的速度之后，Aaron Levie 将视角转向更宏观的商业版图。他认为，软件版图正迎来一场翻天覆地的重构，过去几乎没有软件渗透的传统行业与业务环节，如今都将首次迎来以 AI Agent 为载体的软件形态。在绝大多数知识工作领域，都会诞生专门负责该类工作的 AI Agent。
编程领域正是首个跑通并达到临界规模的样本。从 Replit、Cursor、Windsurf，到 Lovable 与 Cognition，AI 编程工具正以前所未有的速度席卷行业。Aaron Levie 感慨，如果十年前有人预测编程赛道能诞生三四百亿美元的市值，大家一定会嗤之以鼻，因为那时整个赛道几乎只有 GitHub 成功退出。然而今天不仅实现了这一体量，而且“还只是触及了冰山一角”。Aaron Levie 掷地有声地宣称：“我完全有信心为 AI 编程领域承销数千亿美元的市值——这些市值不会流向现有巨头，纯粹是因为这里有太多的接触面，有如此巨大的 TAM（潜在市场空间）需要被满足。”他进一步指出，这一逻辑将平移到生命科学、医疗、法律和咨询等每一个知识密集型行业，新一代 AI Agent 公司不仅会超越传统软件巨头，更将颠覆这些垂直领域的传统外包与专业服务机构。
| 行业领域 | 传统软件与服务模式 | AI Agent 驱动的新范式 |
|:--- |:--- |:--- |
| **软件工程** | GitHub、传统 IDE、外包工程师队伍 | Cursor、Windsurf、Cognition 等背景执行与全自动编程 Agent |
| **法律服务** | 计时收费的律师事务所、律政助理、静态法学检索库 | 具备法务分析、文档起草与尽职调查能力的垂直法律 Agent |
| **医疗与生命科学** | 昂贵的实验室外包、繁重的临床试验管理系统 | 自动化数据解析、药物研发辅助与自动化合规处理 Agent |
| **专业咨询** | 庞大的分析师团队、人工编写汇报 PPT 与商业模型 | 端到端商业调研、数据洞察生成与策略推演 Agent |
### `16:03` 初创公司对抗巨头的结构性优势：背景 Agent 与工程重构
谈及与新一代初创团队交流的体验，Aaron Levie 坦言自己大约有 70% 的时间是在向这群年轻创始人求教。在一年前，创始人对 AI 的运用主要停留在 GitHub Copilot 或是少数尝鲜 Cursor 的阶段，带来的仅仅是 10% 到 20% 的自动补全与替代搜索效率提升；而如今，最顶尖的创始人正在利用具备自主能力的“背景 Agent”（background agents）实现高达 3 倍、5 倍乃至 10 倍的生产力杠杆。
大公司无法直接复刻这种速度的根源在于结构性阻碍。现有成熟企业的基础设施、组织架构和开发规范根本无法支持几十上百个自主背景 Agent 同时在代码库中大刀阔斧地穿梭作业——代码库缺乏足够的文档化支撑、缺乏适配的工具栈，而且一旦数十上百名工程师同时并发运行 Agent，代码冲突将彻底失控。这种混乱恰恰为小团队带来了天然优势：没有历史包袱的微型团队不必担心无休止的冲突，能够最大化释放 Agent 杠杆。Aaron Levie 表示，自己频繁拜访初创企业正是为了观察他们到底如何运转，并思考如何把这些前沿工程方法移植回 Box 的日常运营中。
- 一年前 AI 仅作为补全和搜索工具带来 10%–20% 的效率微调；如今背景 Agent 带来了 3–10 倍的飞跃。
- 成熟巨头的工程管线与组织协作无法承受大量背景 Agent 并发执行所引发的代码冲突与上下文缺失。
- 小团队轻装上阵，天然契合 Agent 优先的开发流，形成了对传统大厂的非对称优势。
### `18:24` 极度杠杆下的创业铁律：合伙人选择与永恒命题
在讨论的收尾阶段，主持人将话题引回到创业最底层的核心要素上：既然两人的微型初创公司借助 Agent 杠杆就能匹敌两千人的组织产能，那么合伙人的选择权重就变得前所未有的关键。主持人发问：如果让现在的 Aaron Levie 去面对当年刚从南加州大学（USC）辍学创业的年轻自己，他会给出什么告诫？
面对是否需要去寻找那些极其冷门、受牌照监管庇护的枯燥赛道的讨论，Aaron Levie 明确回应道：“我认为公司建设的大量核心原则是相对永恒的，它们超越了任何特定的技术浪潮。”他十分赞同主持人的洞察，并特别强调：正是由于 Agent 赋予了个体空前强大的杠杆能力，创始团队与联合创始人的素质不仅没有被技术稀释，反而比以往任何时候都更加关键。在进入具体的项目构思和业务执行之前，最根本的前提依然是选对人并锚定真正的长期价值。
- 无论技术浪潮如何更迭，建立一家伟大企业的核心商业规律依然历久弥新。
- 在 AI 赋能的超高人均产能时代，联合创始人之间的默契、品味与能力上限决定了整座公司的杠杆倍数。
### `20:00` 5000格行业矩阵与垂直智能体的新千亿美元赛道
Aaron Levie 提出了一个极其宏大的预测：未来将会涌现出成百上千家市值达 10 亿、50 亿甚至 100 亿美元以上的 AI 公司。它们的商业模式非常纯粹——抓住经济体系中的某一个具体职能（job function），为该职能打造一个专门的 AI 智能体（Agent），然后将其销售给企业客户、专业服务机构（services firms）或系统集成商。
Levie 构想了一个由所有行业与每个行业内具体岗位交叉构成的“全经济体矩阵”：
> “你可以做一张覆盖所有行业、以及每个行业内每个具体职能的矩阵。那张表上实实在在会有大概 5000 个单元格。”
在这个网格中，比如生命科学行业的销售代表、金融服务业的合规律师等，每一个细分领域都可以用 AI 智能体重新承保（underwrite）。虽然部分格子会受制于潜在市场规模（TAM），有的需要做宽横向覆盖，有的则必须纵向吃透垂直行业，但任何创业者从零起步时，都应该对着这 5000 个格子发问：哪一个行业是我真正了如指掌、能够逆向拆解其工作流程的？
对此，Levie 引用了 Y Combinator 合伙人 Jared Friedman 的一条推文：“去亲自做那份工作，成为执行该工作的专家，然后你就能极具穿透力地做出自动化它的智能体。”他将创业团队的能力拼配分为三个层级：
1. **最佳路径**：创始人过去亲自做过该工作，直接为自己最熟稔的流程打造 Agent；
2. **次佳路径**：拉一位具备该专业背景的人作为联合创始人；
3. **退而求其次**：亲自去对应岗位实习、跟岗（shadowing），整天泡在业务一线。
```mermaid
flowchart TD
accTitle: 垂直领域 AI 智能体构建闭环
accDescr: 展示从选择行业矩阵单元格、深入一线掌握业务流到打造高精度垂直 Agent 的创业路径。
subgraph Matrix["5000格行业矩阵定位"]
A["行业: 生命科学 / 金融 / 法律"] --> B["职能: FDA法规审查 / 案例审查"]
end
subgraph DeepWork["获取行业实操经验"]
C["创始人亲历从业经验"]
D["寻找垂类业务专家联合创始人"]
E["一线跟岗（Shadowing）与业务逆向拆解"]
end
subgraph Product["垂直领域 AI 智能体"]
F["高壁垒专业 Agent\n（如生命科学合规审查智能体）"]
end
Matrix --> DeepWork
DeepWork --> Product
```
---
### `22:20` 劳动成本暴降 90% 引爆的非对称需求（弹性需求释放）
在谈到 TAM（潜在市场规模）时，Levie 强调很多人忽视了劳动力供求关系的弹性。现实经济体系的供求平衡远非完美，大量需求只是被高昂的人力成本生生压抑了：
> “如果有人能以原先十分之一的成本引入某种劳动力，实际上市场对这种劳动力形式的需求会爆发 50 倍。”
以高阶法律审查（Legal Review）为例，传统顶尖律师的费率高达每小时 1,000 美元至 2,000 美元，高昂的价格迫使绝大多数企业放弃了大量潜在诉求。当 AI 智能体将审查成本打至每小时 5 到 10 美元时，市场会突然发现，并不是世界不需要那么多律师，而是以前根本用不起——随之而来的将是全员申请专利、人人都是发明家的大爆发。
在 Box 内部同样存在这种现象。Levie 透露，公司内部有大量工作由于目前人力执行过于低效，在预算堆栈排名（budget stack ranking）中始终被排在最末端，永远得不到立项。但如果这类职能可以由成本只有十分之一的 AI Agent 承接，Box 会立刻投入预算进行采购。
- **雇佣门槛归零**：雇佣一名全职员工需要 10 万至 30 万美元，而引入一个 Agent 可能仅需 1,000 美元；
- **投资回报无上限**：Agent 没有最低准入门槛（floor price），ROI 理论上没有天花板；
- **业务瓶颈转移**：在以前从未设岗的细分盲区部署智能体，企业会意想不到地打通业务堵点，从而促使管理层向该方向追加更多资金。
主持人打趣道，听完这番话后，估计马上会有一大批二十多岁的年轻人涌向 FDA 监管后台充当行政人员，并宣称“是 Aaron Levie 派我来学习业务的”。Levie 顺水推舟提议，加速器（如 Speedrun）甚至可以做一个人才撮合平台：一边是 MIT 辍学、技术过硬的典型 CTO 型创始人，另一边是刚从辉瑞（Pfizer）挖出来的监管合规审查专家，让懂做 Agent 的人与懂工作流的人结对，“直接批量孵化 100 家这样的公司”。
---
### `25:00` 为什么大模型的智商跃升全被企业端吃透？
对话随后转向企业端与消费级（Consumer）市场的对比。主持人指出，Box 最早是以面向普通消费者的云存储起家，随后转向企业服务；以往像 Marc Andreessen 也曾推崇先做消费级产品来磨练产品能力，但本轮 AI 浪潮的重心却极度偏向企业端。
Levie 澄清自己绝不建议创业者“预先策划一场六个月后的转型（pivot）”，因为这违背了转型的本质。他认为消费级市场依然有许多机会，但如果从纯粹的市场规模（TAM）下注，AI 的主战场必然是企业端：
> “AI 从根本上说是一种生产力形式。而生产力通常只在企业端以有意义的方式体现出来；在消费领域，我们并不怎么考虑生产力。”
为了论证这一点，Levie 引用了 Anthropic CEO Dario Amodei 的经典观点：
> “如果 Anthropic 把 AI 模型的水平从化学本科生提升到化学博士研究生，普通消费者在日常使用中根本感觉不到差别。”
在个人与家庭生活场景中，当前的模型能力已经达到了饱和点：
- **日常生活需求已封顶**：Levie 拿自己的母亲开玩笑，称她的日常 AI 诉求已经被完全满足，“哪怕 GPT-5 是 AI 历史上的终点，对她来说也足够了”，因为她绝不会去探究需要博士水平的深度化学或医药研发问题；
- **家庭育儿场景足够用**：给六岁的孩子编儿歌、画插图、讲睡前故事，现有的模型表现已经绰绰有余，不需要故事水平再高深一倍；
- **企业端的极端容错度**：如果将一份严肃的法律合同交给当下的模型跑一遍，依然会挑出 70 处错误。
因此，前沿模型智力从本科跃升至博士所带来的巨大价值红利，将几乎全部被辉瑞、生命科学公司或大型律所等企业级场景所摊销（amortize）。从模型能力的供给端来看，接下来的 AI 演进本质上就是属于企业级的时代；消费级市场的机会将不再来自于底层模型是否具备“博士级智力”，而在于如何通过全新的产品交互形态与包装方式（packaging）去重构用户体验。对话末尾，主持人提及陶哲轩（Terence Tao）对数学工具的评价，并顺势将话题引向下一个核心命题：在产品洞见之外，AI 公司究竟需要怎样的分发渠道（Distribution）洞见。
### `30:00` AI 未曾颠覆分发：科技商业的基础模式依然成立
对谈转向了 AI 时代最具争议的商业命题之一：AI 究竟有没有改变软件的分发（Distribution）逻辑。主持人抛出疑问，AI 确实带来了极其清晰的产品能力与洞见，在产品使用上也展现出自然渗透，但它是否会催生一种截然不同的全新分发范式，还是说我们目前只是在等待平台生态成熟，就像历史上人们从广播时代跨入电视时代那样？
Aaron Levie 给出了一个相当清醒甚至自嘲有些保守的答案。他直言：“我接下来的观点可能会听起来非常平庸，甚至五年后可能被证明是大错特错，但我认为分发并没有发生太大的改变。”（“I'm going to sound really lame and possibly, you know, possibly very wrong in five years from now. I don't think distribution changes that much.”）在 Levie 看来，未来固然会出现诸如 OpenAI App Store、各种“Agent 商店”等新的获客渠道与货架，但这绝对谈不上是整个科技界分发模型的底层颠覆。商业逻辑的铁律依旧未变：做消费级（C端）产品的团队依然必须设法制造病毒式传播（Viral）；做企业级（B端）软件的团队依旧必须脚踏实地建立起成熟的 GTM（Go-To-Market，市场开拓）销售机器；而如果产品恰好处在兼具企业属性、目标客群懂技术且具备自助服务能力的赛道，那么经典路径依旧是 PLG（产品驱动增长）叠加自上而下的 GTM 引擎。归根结底，AI 本身绝不是一种分发维度的颠覆力量。
```mermaid
flowchart TD
accTitle: AI时代的分发模式与商业路径
accDescr: 展示消费端、企业级与混合型产品在AI时代依然沿用的传统分发模式，强调AI并未颠覆分发本质。
subgraph Channels["渠道与呈现形态（出现微调）"]
A["新型分发货架 (如 OpenAI / Agent Store)"]
B["传统流量与社交平台 (X, LinkedIn, Reddit)"]
end
subgraph Models["分发与GTM核心引擎（底层逻辑未变）"]
C["消费端应用 (To C)"] -->|"核心法则"| C1["打造极致病毒式裂变 (Viral)"]
D["企业级软件 (To B)"] -->|"核心法则"| D1["搭建全流程 GTM 销售机器 (Enterprise Sales)"]
E["混合型 / 开发者工具"] -->|"核心法则"| E1["PLG 驱动 + 销售跟进协同 (PLG + GTM Engine)"]
end
Channels -.-> Models
```
- 渠道变迁 vs 模型颠覆：虽然市场上涌现出应用内集成、大模型插件平台或智能体应用商店等新入口，但获客和转化的本质结构并未重构。
- 业务类型的分发规律：
- C 端消费级产品：完全依赖网络效应、裂变与病毒式传播机制。
- B 端重度企业软件：依然依赖大客户直销、安全合规审核、解决方案顾问等重型 GTM 体系。
- 介于二者之间的专业/开发者产品：必须采用“PLG 自助试用 + 销售落地扩大单体合同”的复合打法。
---
### `31:40` “做出来用户就会来”是致命幻觉：构建属于自己的核心受众与分发壁垒
对于当下狂热的 AI 创业者，Aaron Levie 提出了一个极其尖锐的警示：“做出来，他们自然就会来”（If you build it, they will come）在现实商业中绝对是一剂毒药。许多创业者陷入技术至上的自嗨，以为只要打造出了能力逆天的 AI Agent，所有问题就迎刃而解了。Levie 直言不讳地警告：“如果你认为只要做出了这个惊艳的智能体，你就已经解决了自己的生存问题，那你百分之百会走向失败。”（“So if you think that because you built this amazing agent, that now you've solved your problem, like that is like totally you're going to fail.”）在当今的信息过载环境下，市面上有上万家同样打着 AI 旗号的初创公司在争夺眼球，企业客户每天也要处理堆积如山的核心业务优先级，缺乏明确分发策略的初创公司瞬间就会被淹没。
Levie 进一步剖析了当下创业圈对于“流量”的认知误区。很多人把做出了一个播放量极高、点赞拉满的爆款演示视频当作做好了分发，但这不过是一次性的营销噱头。真正坚固的分发，是极其枯燥且日复一日的研磨过程：你必须真正“拥有一群属于你自己的受众”（own an audience）。创业团队需要找到一群对你所解决的问题极度狂热的核心人群，在 LinkedIn、X（前 Twitter）、Reddit 等他们聚集的阵地上建立起持久的社区关系，以“熊抱”（bear hug）的姿态深耕这个垂类行业。在这个基础上，搭配扎实的市场营销、销售团队与公关动作。任何轻视分发的公司，结局只有一种：被分发能力更强、更贴近客户的竞对无情碾压。
对谈尾声，主持人幽默地补充道，看来未来的 AI 创业团队可能会要求每一位员工都必须在所服务的目标垂直行业里“卧底”工作过几个月，就像行业里常说的“只有干过客服的人才知道怎么服务客户”一样。Levie 深表赞同。随后，主持人对 Levie 的深度分享致谢，本期 Office Hours 访谈在对企业 AI 战略与落地现实的清醒探讨中正式落下帷幕。
- 营销噱头 vs 真实分发：一条全网转发的炫酷 Demo 视频不等于分发能力；真正的分发在于拥有持续触达客户阵地的通道，以及高粘性、深参与度的行业社群。
- 拥抱受众（Bear Hug）：团队必须精准锚定所选垂类的受众画像，通过专业内容、社交网络持续渗透，把受众转化为忠实信徒与客户池。
- 忽视分发的代价：在底层模型能力快速商品化的大背景下，产品差异化极易被抹平，如果不在分发上构筑护城河，哪怕技术领先几个月，也会瞬间被拥有更强销售网络与客户关系的竞品碾压淘汰。
## Notable quotes
> `05:00` 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. — Aaron Levie
> `10:00` 现在，拼凑出东西基本上已经商品化了，溢价在于你是否深思熟虑过自己到底想造什么、它是不是一个真正的市场机会、它是不是一个好主意、它看起来美不美、它有没有品味。这些现在成了全新的价值创造形式。 — Aaron Levie
> `11:42` 因为我们面临一个非常现实的问题，那就是：根本不存在什么超级强大、全能的超级 Agent。那行不通。 — Aaron Levie
> `13:40` 我完全有信心为 AI 编程领域承销数千亿美元的市值——这些市值不会流向现有巨头，纯粹是因为这里有太多的接触面，有如此巨大的 TAM（潜在市场空间）需要被满足。 — Aaron Levie
> `18:24` 我认为公司建设的大量核心原则是相对永恒的，它们超越了任何特定的技术浪潮。 — Aaron Levie
> `20:00` 你可以做一张覆盖所有行业、以及每个行业内每个具体职能的矩阵。那张表上实实在在会有大概 5000 个单元格。 — Aaron Levie
> `20:00` 去亲自做那份工作，成为执行该工作的专家，然后你就能极具穿透力地做出自动化它的智能体。 — Jared Friedman
> `22:20` 如果有人能以原先十分之一的成本引入某种劳动力，实际上市场对这种劳动力形式的需求会爆发 50 倍。 — Aaron Levie
## Mentioned
Aaron Levie, Box, Facebook, Dilbert, Claude Code, Replit, Cursor, Windsurf, Lovable, Cognition, GitHub, GitHub Copilot, Y Combinator, Jared Friedman, FDA, Speedrun, Pfizer, Marc Andreessen, Anthropic, Dario Amodei, Terence Tao, OpenAI, LinkedIn, X, Reddit
## Video description
Can a 10-person startup outperform a company with hundreds of employees? Box CEO Aaron Levie argues AI agents are giving small teams a level of leverage that simply wasn't economically possible before, creating a window for startups to outrun incumbents.
In this episode of Speedrun Office Hours, Aaron explains why founders are shifting from doing the work themselves to orchestrating agents, and why that puts a new premium on strategy, taste, and knowing exactly what to build. He makes the case that nearly every industry and job function could support specialized AI agents, potentially creating hundreds of new billion-dollar companies.
They also discuss why enterprise AI may capture much of the next wave of model improvements, how dramatically cheaper AI labor could unlock demand that doesn't exist today, and one thing Aaron thinks AI won't change: distribution. If you build it, they still won't necessarily come.
[Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/cf152414-3a4b-46e0-a3ab-13cdcf42822e/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=cf152414-3a4b-46e0-a3ab-13cdcf42822e&feed=JGE3yC0V)
Source quotes
[5:00] Aaron Levie: 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.
[10:00] Aaron Levie: 现在，拼凑出东西基本上已经商品化了，溢价在于你是否深思熟虑过自己到底想造什么、它是不是一个真正的市场机会、它是不是一个好主意、它看起来美不美、它有没有品味。这些现在成了全新的价值创造形式。
[11:42] Aaron Levie: 因为我们面临一个非常现实的问题，那就是：根本不存在什么超级强大、全能的超级 Agent。那行不通。
[13:40] Aaron Levie: 我完全有信心为 AI 编程领域承销数千亿美元的市值——这些市值不会流向现有巨头，纯粹是因为这里有太多的接触面，有如此巨大的 TAM（潜在市场空间）需要被满足。
[18:24] Aaron Levie: 我认为公司建设的大量核心原则是相对永恒的，它们超越了任何特定的技术浪潮。
[20:00] Aaron Levie: 你可以做一张覆盖所有行业、以及每个行业内每个具体职能的矩阵。那张表上实实在在会有大概 5000 个单元格。
[20:00] Jared Friedman: 去亲自做那份工作，成为执行该工作的专家，然后你就能极具穿透力地做出自动化它的智能体。
[22:20] Aaron Levie: 如果有人能以原先十分之一的成本引入某种劳动力，实际上市场对这种劳动力形式的需求会爆发 50 倍。
Source quotes
[25:00] Aaron Levie: By definition, most AI progress is going to now amortize into enterprise use cases.
[25:00] Aaron Levie: I don't think I would plan a pivot. That kind of breaks the definition of a pivot.
Original transcript
[0:00] Startups can actually outrun these big companies in a way that was just never economically feasible before. You could actually go and compete in almost any category on the dimension of people are missing this whole thing. I really think that there will be hundreds of companies that are all 1 billion market cap, 5 billion market cap, 10 billion market cap and beyond. That basically are just, there's near infinite opportunity for new startups to emerge on the basis of... the ability to deploy resources at a scale that once belonged only to big companies. In this episode of Speedrun Office Hours, Bach's CEO Aaron Levy explains why a 3-, 5-, or 10-person team can increasingly produce the output of a company many times its size. Aaron breaks down how founders are becoming managers of AI agents, why strategy and taste matter more as execution gets cheaper, and why he expects specialized agents to emerge across nearly every industry and job function. part of the startup playbook hasn't changed. Distribution. As Aaron puts it, if you build it, they will not necessarily come. Aaron, welcome. Thanks for joining us. Thanks for having me. It's kind of interesting that you're actually still a young guy, only a few years older than me and probably similar to Andrew, but you've been running Box since 2005, which is almost exactly half your life. So if you could compare the world of startups now to the world as it was back then, what do you think about the startup experience has changed? My biggest takeaway actually is I feel really old. So I appreciate, um, I appreciate that you're using your age as the anchor for what young is. But, um, uh, but, but I think unfortunately, uh, I, I tend to feel like very old now. Um, and, uh, I mean, I am like so impressed and excited by so many of these young founders right now. And I'm You know, what's interesting is that, In 2005, so we dropped out of college, our middle of our junior year, and in 2005, it was fairly rare to drop out and kind of be 20 doing a startup. There were a few dozen of us in Silicon Valley at that time. I think... Prompted probably mostly by like Facebook had made dropping out like a thing, but it was still a relatively rare event. And what is so fascinating now is the amount of now 20 year olds that are dropping out. It's it's like it is just like now a standard playbook for these kids. And and so I think it's you know, it's I think it's super exciting how. just how many young people are getting started and jumping right into building startups. So I do tend to feel pretty old. But yeah, the playbook seems incredibly different. I mean, the way that startups are being run right now, is the most different than any other period of technological change that I've ever seen. Where if you looked at how we ran Box in 2005 and you compared us, to a startup 10 years prior. So let's say you took the 1995 company and the 2005 version of a startup. It actually wasn't that different. At least in 2005, it was a little bit right before the cloud wave. We still went to data centers. We racked servers. We coded morning and night. We were a startup. But if you looked at how we ran the company in 2005 compared to 1995, it would almost look exactly the same. The way that every 1995 startup got started looked exactly like how we got started and ran the business. If I compare the 2025 startup to us in 2005, or let's say the 2015 startup, so if you just use 10-year increments, It is the most different that I've ever seen in the history of my adult life and doing technology now for 27 years, let's say. And it's simply because of this AI first orientation. where you can be a three, five, 10-person team and have the output of a 50 or 100 or 200-person team And that just changes everything about how you run these businesses. So I'm wildly impressed by these startups. I also think some of them are just totally crazy and unhinged. But it's fun to watch and fun to learn from. But it is the most different that I've ever seen how to build a company today. On that point, it's kind of interesting. In the case of where you and your co-founder, you guys dropped out, if you had the ability to force multiply with AI agents, for example. I'm curious what you're seeing. I think you wrote about this recently, how some of these startups are, in particular, I think doing things in the background and they're using AI agents in interesting ways. What's been the most craziest thing?
[5:00] for her seat so far. Well, if you just take the basic premise and then draw it out. Yeah. When we started Box... Our rate of progress as a startup was 100% correlated to how many hours we could be on a keyboard and type things. That was the... The only point of leverage we had was how fast we could type text into a keyboard. and we could be typing code, we could be typing marketing assets, we could be typing blog posts, - Right, right. - We were managing servers, But the four of us, we all kind of dropped out. We lived and worked in this renovated garage, kind of live-work kind of thing. And that was the only speed at which we could execute was how fast we could work on computers. The startups that I'm talking to now... It's, This is not exclusively the case, but let's say it's increasingly the case. It seems like there's now a new sense where the startup founders and the early employees, and I think this basically, you know, sort of scales to every industry over time. They're more managers of work. And the work is being done by agents. And so they're no longer rate limited by how quickly they can type and how quickly they can write code and how quickly they can come up with marketing assets. They are deploying agents to go and do a lot of that work. And I was talking to a startup founder recently that, claims at least, and I basically believe it, he's just reviewing thousands of lines of code a day, that the agent did, And the ability to be an editor or a reviewer, is like 10x the leverage of the person who actually has to create the thing. And so if you kind of have that as your mental model, where this person now is the editor, orchestrator, reviewer of code, And they can go on and deploy multiple agents, so they're no longer rate-limited by, again, like one terminal, one user interface that they can interact in. You can only... type in one IDE at a time as an individual kind of programmer, but I can review I can have lots of agents now running in the background, and I can go review their work, when they're all done, that's just a totally new form of leverage. And so now the bottleneck is just reviewing code, orchestrating it, integrating it, kicking off agents. The bottleneck is not just how much code have you written? So I'm mostly seeing it on the engineering side. So the big benefit is it's giving startups a new form of leverage that never existed in the startup world. Usually you have this deficiency as a startup, which is you have two or three people and you're attacking some big incumbent that is 100 times more people. And you can basically move faster than that incumbent, but that incumbent can always deploy more resources. While we are increasingly neutralizing that delta because of agents. We're not fully there because agents make mistakes. It's not a panacea. A lot of people describe it like a slot machine. You have to keep trying over and over again and seeing what the agent comes back with. But you can kind of squint and sort of see what this could turn into, which is imagine a world where the agent could do a few thousand lines of code that was really high quality and adhered to your code base standards. and it really worked, and you reviewed the code, and you only changed a few things, Well, that would mean that you do have now 10x the engineering capacity as a startup, which means you could actually go and compete. in almost any category, on the dimension of scale and speed, which would change everything about competitive dynamics. And then as an interesting just asterisk, The big companies, I find it very, very hard to actually go and equivalently do that. because they have a lot of workflows, a lot of processes. that are so wired into how they already operate that they won't be able to make the change fast enough. 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. So again, starting in code, but I think the most kind of wired in companies are figuring out how to do this in marketing. They're going to figure out how to do it in legal work. They're going to figure out how to do it in outbound sales. So that will be a pretty interesting evolution of the start formation. I guess the thing I'm wondering about Aaron is that we've always talked about before that like Sometimes everyone wants to work at the speed of strategy. And they want to be able to work at the speed of creativity and thought. Nobody wants to spend all their time writing marketing copy, unless you're Silicon Valley's favorite internet comedian like Aaron Levy. But I mean... I do think that there is this question, though, that if you are now can be more strategic, though, the decisions you make actually matter quite a bit. Yeah. Like you could really like you might really like mess up. And then luckily it doesn't cost that much, but you're still going to mess up. And in fact, like really the speed of creativity and and sort of the chessboard actually matters more. Yeah, I think, you know, one thing that that is interesting is is so that the tendency appears to be.
[10:00] that the way you're building software now, at least in these startups, is you get really, really clear prompts or specs. You get very, very clear ways of working with agents to understand your code base, to build sub-agents at the right parts of your code base. And there's an increasing view of this sort of spec-driven development, which is you write very, very in-depth specifications about what you're doing. Sort of the idea of the product engineer, right? Like the product person. And actually, you know, the pointy haired Dilbert product manager now truly actually has the power to do their own stuff. You wait long enough and they become very, very powerful. But yeah, so it turns out if you're really good at writing a spec and you're really good at knowing and having clarity about what you want to build, that's actually extreme leverage now. So there's a huge premium on knowing, on actually being very good at strategy, being very good at like what is the market opportunity for what you're building. and not just, previously there used to be probably a premium on just how quickly can I hack something together? Well now actually because the hacking something together is fairly commoditized, the premium is have you thought through what you actually trying to build is a real market opportunity for it is it a good idea Does it look good Does it have taste Those are now the new forms of value creation And so I do think that the companies that are really religious about getting that right will be the ones that get ahead in the future. I guess the question behind this is, okay, so let's assume that if you're a startup, you effectively have all these tools at your disposal. You could obviously adopt one of the bigger folks, right? And then the issue with that is you've argue around this idea of sort of AI context rot, which basically the way to solve that is to actually have these sort of point solutions potentially, or great sort of well-designated agents, as opposed to having the single sort of ultimate, the ultimate fighter to actually have a team of the ventures to go and help you in the world of AI. I guess if you're that startup, right, and you're trying to figure out what's the agent architecture that you kind of ship with first, what's the fast way to get off the rails, it's one of these things where, you know, it's at what point do you just give up everything? And at what point do you start filling out, you know, do you just start with like a basic toolkit, for example? Well, the point I was making in that context is not so much about the number of tools you use as much as the need for specialization in some virtual or very tangible way And so like for instance Cloud Code has created a way to have that virtual specialization by just letting you go and create subagents that own a particular part of your code base or a particular part of your workflow. And that's sort of how you resolve the context rot issue within the Cloud Code kind of paradigm. So it's not so much that you want a tool for everything. It's that we have a very real issue, which is this, you know, like super powerful, super agent doesn't exist. It doesn't work. And so you do have to have some degree of specialization and that specialization can come, you know, in a variety of packages. I think there's like, there's a YouTube video on every, every way to solve this. You ask, you know, three startups or five startups, you're going to get three or five different answers. And it's totally fine because this is sort of, this is the stage where everybody's kind of figuring out the best practices and it's all emerging right now. But I don't have a, I don't, I can't give any kind of general principle on that because it's just different by team, by company, by what tool you choose. But it would seem that at the very least your perspective is there is definitely not like, this is not just going to be able to go and benefit incumbents. There's this unique opportunity. If you could be one of those specialized tools as a startup you in a really really good spot You can actually remake the whole area of the stack Yeah okay So now if I flip it a little bit because I was mostly answering the question through the lens of generically speaking the fact that AI startups can move 10 times faster than an incumbent, that's fantastic. Now, if you think about it, okay, what does that mean in terms of what you should build? How should you compete in this market? I think there's infinite, that's obviously a little hyperbolic, But there's near infinite opportunity for new startups to emerge on the basis of we are going to see a complete shift in the software landscape because so many industries and parts of business that previously didn't have software now will have software in the form of AI agents for the first time ever. And if I look at most domains of knowledge work, we'll have some kind of agent that is associated with that type of knowledge work. Incumbents will take some categories, but there'll be startup opportunities in every single category for basically building agents that do knowledge work. And the first ones that have reached some degree of critical mass are coding agents. And there's been very clear startup opportunities in the form of replet,
[15:00] Cursor, Windsurf, lovable, cognition. Holy crap, if we had said 10 years ago that there'd be $20, $30, $40 billion of market cap in the coding space... And we would have been like, well, GitHub is the only thing that's ever really even exited in the coding space. How could there be 30, 40, 50 billion dollars within just a three year period that wouldn't sound plausible. And yet that's not only happening, but it's actually just scratching the surface. I'm perfectly comfortable underwriting a couple hundred billion dollars of market cap. in the AI coding space that don't go to incumbents, purely because there's just so much surface area, so much TAM that needs to exist. Okay, now do that for every single field. Do that for life sciences, do that for healthcare, do that for legal work, do that for consulting. There is going to be new AI agent companies in every single one of those spaces. And not only will they outrun the traditional software incumbent companies, they're going to outrun the traditional services companies in those spaces as well. And it's just a completely new environment for startups to emerge. You said something at the beginning. The beginning of this conversation that I thought was interesting, which was when you talk to founders, you feel like you're you're doing the learning about 70 percent of the time. I was curious if there's anything that comes to mind, like what are you learning from kind of the new generation of founders when you do speak with them? Yeah, I think it's mostly, I mostly just have to rethink the leverage that you get from agents in every conversation. So if I took an example conversation I would have had with a founder a year ago. And you say, hey, how are you using AI? And again, the standard at that point would have been GitHub Copilot, maybe a few users on cursor, they would say, yeah, I'm getting 10 or 20% productivity gains. It's really helping me with type ahead, autocomplete functionality. It's like replaced me needing to Google. So when I have a coding question, now if you ask the most advanced startup founders, how are they using AI, it's back to this point. It's not a 20-30% form of leverage. It's possibly a 3-5-10x leverage. form of leverage because of this ability to have background agents. But the problem is, again, in your existing teams, and your existing organizations, your existing way of doing engineering is not wired up for kicking off lots of background agents to go and execute on your code base. Your code base isn't designed in that way. You don't have the documentation in the code base in a lot of times to support agents to go and operate on the code base. You don't probably have the tool stack to be able to do that. You haven't done the engineering methodology change to enable dozens or hundreds of people to go do all this at once. It actually is something that benefits a small team where they're not going to have all these conflicts happening at all time across all these agents. And so this is something that naturally is advantaged for the smaller startup to be able to move quickly on. And so when I'm talking to startups, I'm mostly just learning, how are you doing that? Like, and then trying to figure out like, what's my way of bridging our company to do that as quickly as possible so we can take, you know, a lot of the benefit of these learnings and port them into how we run Vox. That makes so much sense. All right, so let's, you know, circling back to the first of our very first question, I'm curious how you're thinking about, if any, important things you think folks need to know. Obviously, we've talked about leverage, agents, the ability to basically be, you're a two-person startup. So actually, choosing your founder really matters a lot more. Effectively, that's equal to 2,000 people. But I mean, thinking through that paradigm now, if you were you know, Picture... Aaron from dropping out now. Right? what would you, what would you, you, you tell Aaron now? Yeah. Um, We've already talked a lot about it, but what are the more basic building block things? Maybe, does anything change there? I think a lot of the core principles of company building are relatively timeless, and they transcend any particular technology wave. To your exact point, team probably matters more than ever before because of the leverage on your co-founders or co-founding team. I think that's a great insight. That was yours. I just would underscore that point. for example. Some folks have said, try to find the most highly specialized, really boring market on the planet to find regulatory capture, maybe. Or you're like, forget about that. That's not going to work for you, man. No, no, no. So, okay. So, So there's some things that are downstream from you have the idea, but if you want to go all the way up to you don't have the idea yet, then I'll be able to start there. I think...
[20:00] I really think that there will be hundreds of companies that are, are all... 1 billion market cap, 5 billion market cap, 10 billion market cap, and beyond. that basically are just, they're taking a job function in the economy, and they're building an agent for that job function, and they're selling that agent to corporations, to services firms, to system integrators, to everybody, and And I think you could literally do, maybe Andreessen has one of these, you could literally do a matrix that, That is every industry and every line of every job function per industry. And it would be like, it would be literally 5,000 cells. And it would be like, There's sales reps in life sciences, There's lawyers in financial services, And some of the markets, you know, probably will be TAM constrained, but some will be much bigger. Some need to be kind of like, okay, we have to get the whole horizontal. Some will have to be wide to get the whole vertical. But I think you could very easily go through the entire economy and, and underwrite AI agents for any one of those slots in the economy, And, you know, if I had to literally start from scratch, I'd probably be looking at that matrix and saying, OK, is this is there an industry I know extremely well? And that I can go in and kind of really reverse engineer, like, what does the work look like in that industry and build an agent for that? Jared Friedman had this great tweet. He's a Y Combinator and he basically said, "Go do the job." of what you want the agent to go and automate. and be like the expert in how to do that job. And then you're going to be very potent at being able to go and create an agent to do that job, right? And so, you know, probably like the best case scenario is you've already done that job in the past and you just go and build an agent for the work you already know how to do. Second best scenario, have a co-founder that has done that. Third best scenario, go literally find a way to do that, shadow somebody, do it all day long. And then the opportunity is, can you make an agent that is like the life sciences FDA regulatory review agent? Can you go do the, I am the legal review case agent? Right. And there's going to be, again, there's going to be a lot of flame outs in this space that are, again, they just didn't nail the scale. It wasn't the right market fit. The TAM wasn't big enough. There'll be lots of that. But then there's going to be a where it'll turn out that there's more categories than we think, where right now the supply-demand equilibrium in the economy is actually not sort of as perfect as we would have imagined, where 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. That's right. And so you'll all of a sudden be like, okay, I'm a legal review agent, and it turns out that traditionally legal review costs $2,000 per hour or $1,000 per hour for high-end work. Now the agent can do that for $5 or $10 an hour. And it just turns out that we didn't actually have as many lawyers as the world needed. It was just that they were so expensive, so nobody could actually afford them. And then boom, there's an explosion of this new category of everybody's filing patents now everybody is going to be an inventor of some sort yeah I can name a bunch of use cases even internally. I'm not going to give too many startup ideas out there, but I can name a bunch of ideas internally where... If the cost of the particular function was a tenth of the price, We would spend we would have many more kind of full-time people equivalent of that function. It's just too inefficient right now. And so it is always at the bottom of our stack ranking budget planning list. So we never get to it. But if it was an AI agent that could be compressed... At a tenth of the price, then we actually would be spending on that agent. And amazing. Because there's no ceiling. There's no ceiling to the ROI potentially, right? There's no ceiling to the ROI, and there's no minimum floor price to get started. And that's the, like, people are missing this whole thing. Like, to hire a person costs, $100,000, $200,000, $300,000 or whatever, depending on the function. To bring in an agent might cost $1,000. And so you will start using them in use cases that you never hired the person for. And that will then, all of a sudden, it'll kind of reduce the bottlenecks in your business in ways that you didn't anticipate, which then will cause you to actually go and fund those areas even more. So I think there's a lot of opportunity that will kind of look like that. But again, for an industry, for a job function, and I think there's endless opportunity in that general space. - I look forward to after this, there's a, all of a sudden, there's this giant surplus of young 20-somethings who end up as like admins in FDA regulated back offices. Where did all these people come from? They're like, "Oh, Aaron sent me." Okay, I'm just, they're just here. They're really excited to go and work here.
[25:00] or industry. I mean, what we probably need is like, we need this like pairing mechanism between, maybe you guys start it. It's like, you're a strong, you're like very clearly central casting CTO, you know, founder, you dropped out at MIT. And then you pair them with somebody who just like, you pluck them out of Pfizer. And you're like, you know how to do regulatory review processes. We're just going to match. And then there's just, it's just a matchmaking service between the person that knows how to build agents and the experts in the workflow. and we just go create 100 of these. I'm curious to hear about this, right? Because Box started out in the consumer world. Then you guys went to enterprise. And in the past, even Mark himself over at, you know, Ed Rees Horowitz once said, he's a really great founder to actually start out in consumer just because it's so hard to do consumer. You have to build products. It's very difficult to do. But in this wave, obviously with AI agents, it's way more focused on enterprise. And we're sort of making a tack where we kind of push toward basically enterprise. How do you think about consumer? Or do you still think that's a proving ground? Or is it one of those things where it's like, hey, that's just a completely different world that no one knows how that going to change as a revolver AI I think there plenty of consumer opportunities I don think I would pre a pivot That was not planned Aaron It kind of breaks the definition of a pivot. Yeah, that's right. In six months from now, we are going to pivot the company, but we're going to first start in a different state. I would say that it's probably still very good advice if people do need to pivot, but I think there's going to be plenty of consumer opportunities. And I'm partial to the enterprise because it's a space that I spend a lot of time in. I think by TAM, I would largely bet on AI being a breakthrough for the enterprise market by just sheer size of the TAM. Simply because AI is fundamentally a form of productivity. And productivity generally shows up in a meaningful way in the enterprise. In the consumer space, it's not a thing that we think about as much. So I just think most of the use cases tend to lean enterprise. And it's interesting, Dario from Anthropic had a point on a recent podcast that I think is the most salient way to think about it, which is if Anthropic improves the AI model from basically being like an undergrad in chemistry to a grad student in chemistry the consumer won notice in cloud There no question you're going to ask about chemistry in your personal life that is going to require that PhD level. I mean, barring you have some deep healthcare related issue. But the PhD jump is going to have massive gains for Pfizer and the life sciences industry. And so just if you think about where we are in the kind of progress of AI, I think most of my, I'm going to just pick on her just because it kind of makes the point. My mom's AI use cases, they're fully solved. We're done. She would be fine if GPT-5 was just the last thing that AI ever produced because it's already solving every problem. Because she's not going to know more chemistry questions. She's done with her chemistry questions. She's not trying to create a new life sciences drug. Even in my personal life like for parenting let say I think AI could stop right now and it would already have been a huge net positive for me We make songs for the kids We draw things We get stories made I don need stories to be any better for my six We have reached the full saturation of capability that is needed in most of my personal life. But again, you go to a law firm and you're like, hey, I ran this contract through an AI model. and it's like, well, there's like 70 mistakes. So by definition, and back to the original point, by definition, most AI progress is going to now amortize into enterprise use cases. So from here on out, it's an enterprise kind of industry with, again, plenty of opportunity for starting consumer companies, to be clear, but it's not because we needed the PhD thing. It's because we needed new packaging in a consumer app that sort of brought this experience together in a better or a new way. I mean, this is the sort of test that I always check. We have Terry Tao, the Australian mathematician, says something, then I'm just like, oh, things are going really well. He's gonna get us warp engines very soon. So I guess, curious how you think about the, we talk about this a little bit, which is that you should have product insights, but in addition to product insights, you sort of need distribution insights. And it doesn't seem like, this is like a big debate where
[30:00] it'll say, AI has provided some clear product insights. But it hasn't changed the world of distribution anyway, but we obviously see some natural adoption of AI tools. Um, But I'm curious how you think that might evolve. Is it because we're just waiting for things to platform, and then we'll have distribution insights in that regard? Or do you think that actually, it's a whole new ballgame, we're thinking about it like radio and television right now? I am going to take the, I'm going to sound really lame and possibly, you know, possibly very wrong in five years from now. I don't think distribution changes that much. - Ooh, interesting, okay. - I think that the, you know, there'll be new channels, like you'll be like, "I'd like to show up in the OpenAI App Store," you know, Asian store, like there'll be new ways that you think about getting distributed. But I don't think this is a fundamental transformation or disruption of the distribution model of tech. If you're a consumer, you better figure out how to go viral. If you're an enterprise, you better figure out how to do, you know, build a go-to-market machine. Um, if you're, uh, if you're fortunate and you're kind of like, and you're like enterprise-y, but your audience is somewhat technically savvy or... or able to do self-serve, then you'll do PLG plus a go-to-market engine. But I don't think we're going to see a new invention of go-to-market technology. because of AI. I don't think AI is a go-to-market related disruption model. Maybe there's an important corollary actually, and this might be appreciated by AI startups. Bye. If you build it, they will not come. So if you think that because you built this amazing agent, that now you've solved your problem, like that is like totally you're going to fail. Like, you are going to be drowned out by 10,000 other startups and priorities that companies deal with all day long. So if you do not think about distribution, then it just won't work. And I think we might have a little bit of an over-orientation toward the, hey, I've made this really viral video and everybody saw it and I got all the likes. And that's not distribution. That's like a cool one-time marketing stunt. Like distribution is just the grinding of like, The way I think about it, and I think AI affords this quite effectively, is that it's You have to own an audience. There has to be a universe of people that are just so passionate about the thing that you're building. You need to figure out how to just build, continue to keep access to that relationship. So you better own your community via LinkedIn and X and Reddit and wherever those people are. And you need to find a way to just bear hug them as an industry or as a cohort. And then that's going to be marketing, that's going to be salespeople, that's going to be PR, all the things. And the companies that discount that, you're just waiting to be rolled over by somebody who has better distribution and just says, okay, I'm going to the customers. I look forward to every founder basically requiring every employee to have worked in that industry as a blah, blah, blah person. And then being like, oh, I worked in that industry for five, three, four months. Just like when people say like, you have to have worked in customer service to know how to deal with customers. Exactly. All right, Aaron, thanks so much for doing this. That's a great one to end on. So we really appreciate you for joining us. Thanks, guys. Thanks again for listening, and I'll see you in the next episode. Disclosures. you
