Aaron Levie | The a16z Show Summary

Podcast: The a16z Show | Participants: Aaron Levie (Box CEO), Erik Torenberg (a16z GP), Martin Casado (a16z GP), Steve Sinofsky (a16z Board Partner)

1. Key Themes

Agents Will Force a Complete Rebuild of Enterprise Software Architecture

The conversation returns repeatedly to a single, structural insight: if agents outnumber human users by 100x or 1,000x, the entire software stack must be re-engineered to serve agents as primary consumers — not humans. This has massive implications for every SaaS vendor alive today.

"If you have a hundred or a thousand times more agents than people, then your software has to be built for agents." — Aaron Levie 00:00:28

"Your business performance will correlate to how well your agents can get access to the information they need to do their work. And so thus, your enterprise IT stack has to be set up in such a way to support that." — Aaron Levie 00:31:54

The practical implication: the new moat for software companies is not UI polish or brand — it's API quality, agent identity management, and access control infrastructure.

"The game is, can you build really, really high quality APIs? Can you have a way of monetizing that? Do you have a way of handling the identities and all of the access controls for agents? And that becomes the new problem you have to solve if you're building a software company." — Aaron Levie 00:32:24

AI Capability Diffusion Will Be Much Slower Than Silicon Valley Expects

Both Levie and Casado converge on a sobering point that runs counter to the dominant Valley narrative: the gap between what's technically possible and what's deployable at enterprise scale is enormous, and will persist for years. The reasons are structural, not just technical.

"A lot of this is actually why the diffusion of AI capability is going to take longer than people in Silicon Valley realize because what's happening is like we see startups that can start from the ground up without any of the risks that we're talking about because they have nothing to blow up." — Aaron Levie 00:27:50

"It's just absurd to think you're going to vibe code your way to like SAP. All of that domain knowledge, it's not just represented in some well-orchestrated data layer." — Martin Casado 00:00:00

Casado frames this using the analogy of his cousin learning spreadsheets — the abstraction layer always moves up, but always slower than the technologists expect:

"I feel like when I talk to people trying to do stuff that we're right, I feel like I'm at Thanksgiving dinner talking to my cousin six months in her job. When I'm using a spreadsheet already and I'm like, I don't know why this is so hard. You should just use one. And then two years later, she's doing it." — Martin Casado 00:08:18

The Agent Security & Identity Problem Is Fundamentally Unsolved — and Underappreciated

The deepest and most technically substantive thread of the conversation is about why agents cannot simply be treated as human users — and why the current model of "give the agent a credit card and its own account" breaks down at enterprise scale. This is a largely unsolved problem with no clean answer yet.

"The agent, you have all the liability of whatever they're doing. You do have complete oversight and you're probably going to need to have that complete oversight. They have no right to privacy." — Aaron Levie 00:19:46

"If I know your new agent's email address and I email it like it's an assistant... I can social engineer it 10 times easier than a human." — Aaron Levie 00:21:52

"The threat vectors are going to be way more sophisticated... you can't just assume that the agent acts like a human does today because it's going to be the fastest, most thoughtful, craziest-ass human that ever existed trying to actually leak the information because it got injected in some way." — Martin Casado 00:25:30

2. Contrarian Perspectives

Marketing to Agents via Better Interfaces Is the Wrong Strategy — Substance Wins

The prevailing industry advice is to "build for agents" by improving documentation, APIs, and marketing surfaces. Torenberg pushes back hard, arguing agents are actually better than humans at cutting through interface noise to find the best underlying system.

"People in the abstract say things like, now you're marketing to agents, you're like an API, you've got a good idea. I actually think that's almost exactly wrong... Agents are very, very good at picking the right back end for whatever they're doing. They're like, the cost parameters of this, the durability of that... they actually have the collective wisdom of our experience using these platforms." — Erik Torenberg 00:33:54

"I think as an industry, we're so focused on these interfaces... But really, I think that we're going to be pushed to actually build better systems. And that's what's going to be chosen." — Erik Torenberg 00:34:46

The implication: agents will select backends based on actual performance, reliability, and cost — making the quality of underlying infrastructure a meritocratic competition for the first time.

Enterprise Software Layers Will Not Collapse — They Will Persist and Multiply

Against the "prompting to machine code" collapse narrative popular with figures like Elon Musk, Torenberg and Casado argue history is unambiguously clear: layers never disappear. They persist because they encode organizational and compatibility logic, not just technical logic.

"The history of systems is layers never go away. They just get layered... because a lot of the layers are actually more of like organizational boundaries or state boundaries or compatibility. They stay for compatibility." — Erik Torenberg 00:38:26

"This is not, it also hasn't been, not tried before. Like if you were to look at an ERP system from first principles, you know, well, in 1970... today you would start from a different set of assumptions... But then it would still only last like 10 years until you thought, wow, that was a broken decision." — Martin Casado 00:39:22

Wall Street's AI Revenue Models Are Wrong by an Order of Magnitude — Upward

This is perhaps the most direct contrarian call in the episode. Casado argues with conviction that analysts modeling AI revenue are making the same mistake they made with PCs and cloud — treating it as a linear substitution rather than a multiplicative expansion.

"The biggest problem right now is everybody is trying to figure out the economics of all of this. When they're off by at least an order of magnitude on how big the opportunity is." — Martin Casado 00:00:19

"People viewed PCs as a finite market because they just viewed the consumption of MIPS as some finite thing and they didn't think what would happen if we put all those MIPS on every desktop... And it turns out that was like a really good idea — and the same thing happened with the cloud... nobody went, oh, people are going to use a thousand times as much of the resource if we move it there." — Martin Casado 00:43:52

He supports this with live portfolio evidence from Torenberg:

"Every single one of them has gone asymptotic in the last six months... It just turns out there's so much more software being written now than ever has been before." — Erik Torenberg 00:45:22

The Anthropic Growth Marketer Example Is NOT Representative of the Future of Work

Casado pushes back firmly on using the viral "one person + Claude Code = 10-person team" story as a template, pointing out it only works under unrealistically favorable conditions — infinite demand, hot product, no competition.

"Using the anthropic growth person as an example — that is a job. That is the rest of work... When demand is infinite and frankly supply is infinite, this is not a difficult job... Be instead the $600 PC marketing person and see how you can do against the NEO. That's a real job." — Martin Casado 00:05:57

Agents Will Inadvertently Create Rogue Systems of Record — Inside Companies

Casado raises an underappreciated risk: agents spinning up their own de facto data layers inside the "middleware" zone that IT historically ignores — repeating the exact pattern of uncontrolled file shares and marketing databases of the 90s and 2000s.

"The agents themselves will spin up what becomes like a de facto new system of record... in what the IT people think of as some middleware, end user BS area... like in a sense, the macros end up running the corporation." — Martin Casado 00:36:52

3. Companies Identified

Box Cloud content management and file storage platform for enterprises. Mentioned throughout as the primary case study for how an enterprise SaaS company is actively building for the agent era — investing in CLI access, agent-facing APIs, MCP integration, and agent identity management.

"We just rolled out the official box CLI... you give Claude Code the box CLI and you can now interact with your entire box system via natural language. And you get the horsepower of Opus 4, 6 being the orchestrator of doing a bunch of operations." — Aaron Levie 00:16:02

Anthropic AI safety company, maker of the Claude model family. Cited multiple times as the platform powering advanced agent workflows, including the viral growth marketer example and Box's own CLI integration using Claude/Opus.

"He was using Cloud Code at the time to basically more or less automate what maybe five or 10 people would have done." — Aaron Levie 00:05:04

SAP Enterprise ERP software giant. Used as the canonical example of legacy software that is not going anywhere — representing the immovability of domain-encoded enterprise systems and the ceiling on AI diffusion.

"It's just absurd to think you're going to vibe code your way to like SAP. All of that domain knowledge, it's not just represented in some well-orchestrated data layer." — Martin Casado 00:29:59

Salesforce CRM platform. Cited as the archetypal example of a market that appeared fixed ($2B CRM industry) but exploded when distribution friction collapsed — used as the template for what AI economics will do.

"The CRM business... was two billion a year... When if you could just get salespeople to sign up individually, they all will sign up with no friction. And that is exactly what's going to happen with AI." — Martin Casado 00:45:20

Workday Enterprise HR and finance SaaS. Referenced as an example of a legacy SaaS vendor facing the structural tension between their current intelligence/UI-based monetization model and agents that only want raw data access.

"How does Workday charge a penny for every HR record it pulls? We'll figure that out." — Aaron Levie 00:32:52

Perplexity AI-native search and research platform. Named as one of the emerging "super app" paradigm players where agents can transact and retrieve information on demand.

"That's the cloud co-work phenomenon. That's whatever OpenAI is kind of cooking up with the super app, perplexity computer, et cetera." — Aaron Levie 00:02:43

4. People Identified

Aaron Levie CEO and co-founder of Box. Stands out in this conversation as one of the most practically grounded AI-era enterprise thinkers — he's not theorizing but actively deploying agent infrastructure at scale, grappling in real time with identity, access control, compute budgeting, and monetization challenges.

"We actually have an agent that we're working on where it just makes a determination whether it should use an existing skill... or it should write code to solve that problem. And its ability to do any one of those three at any moment ends up being incredibly useful." — Aaron Levie 00:10:09

Martin Casado General Partner at a16z, former founder of Nicira (acquired by VMware for $1.26B). The most contrarian voice in the room — consistently grounding the conversation in the realities of enterprise IT, the limits of algorithmic thinking, and historical technology diffusion patterns. His skepticism is well-substantiated and consistently backed by specific analogies and lived experience.

"The diffusion of AI capability is going to take longer than people in Silicon Valley realize." — Martin Casado 00:00:00

Mark Benioff CEO of Salesforce. Cited as the paradigmatic example of an entrepreneur who correctly identified that a "fixed" market was actually unlimited once friction was removed — held up as the mental model for how to think about AI economics.

"Mark was just blazing the trail, which was like the CRM business... was two billion a year... When if you could just get salespeople to sign up individually, they all will sign up with no friction." — Martin Casado 00:45:20

Bill Gates (and Paul Allen) Founders of Microsoft. Cited as the canonical example of the person who saw software as a separate, scalable business when everyone else assumed it came bundled with hardware MIPS — the original "order of magnitude" revenue insight.

"People thought software just came with the MIPS and nobody thought, oh, well, they'll just sell the software. One guy did. And it turns out that was like a really good idea." — Martin Casado 00:43:52

5. Operating Insights

The Engineering Compute Budget Is Now a Strategic CFO-Level Decision, Not a DevOps Detail

Levie flags that allocating engineering expense to tokens is a fundamentally new financial management challenge — one that will define EPS at public technology companies and cannot be treated as a line item optimization.

"The engineering compute budget conversation is going to be the most wild one in the next couple of years... R&D is somewhere between 14 to 30% of revenue of any public technology company. The difference between compute being 2x the cost of your engineering team or 3% more is like that's all your EPS." — Aaron Levie [00:00:13 / 00:48:24]

Practical implication: operators need to proactively build frameworks for token budgeting now — including decisions about parallel experimentation vs. efficiency, and which workflows justify long-running agents vs. single-shot prompts.

Build for Read-Only Agent Access First — Write Access Comes Much Later

Levie offers a practical sequencing principle for enterprise AI deployment: a read-only layer of agent access is achievable now and safer, while write/execute access requires unsolved identity and security infrastructure.

"I think we have a read-only version of this for a number of years before... where N is very large." — Aaron Levie 00:15:49

This maps directly to a go-to-market and product sequencing strategy: ship agent-readable APIs and data layers immediately; defer autonomous write/execute capabilities until identity and audit infrastructure catches up.

"Knowledge Worker Services Companies" Built from Scratch with Agents Are the Near-Term Disruption Vector

Levie identifies a specific, actionable category of new company that is uniquely advantaged right now: services businesses (marketing agencies, legal, engineering consulting, architecture/design) built from zero with agents — no legacy access controls, no information silos, full context availability.

"If I could start a marketing agency or consultant, engineering consulting company... you could kind of build your company pretty differently if you had no constraints... I can give the agent just all the context it needs to do its work... I do think that will be relatively disruptive for some time until the bigger incumbents can kind of get out of the way." — Aaron Levie 00:41:20

6. Overlooked Insights

Micropayments for Data Are Finally Viable — Agents Remove the Friction That Killed Every Previous Attempt

Buried in a brief exchange is what may be one of the most commercially significant structural changes in internet economics: agents will unlock an entirely new category of transactable data assets that have been economically stranded for decades because humans would never pay small amounts for them.

Every micropayment scheme from the 1990s onward failed because humans experience transaction friction psychologically — they won't pay $0.05 for an article or $3 for a medical study. Agents have no such friction. This means vast libraries of paywalled, underutilized data suddenly have a buyer.

"The agent doesn't care about the friction of a small transaction. It's the first time that you can have resources behind a paywall that something will actually be willing to pay for that resource... you could go get medical research for some deep research task they're doing and I'll pay $3 for that and the agent is able to go and transact. It kind of opens up a whole new world of business models for the internet." — Aaron Levie 00:46:41

This point passed with minimal discussion, but the investment implication is large: data publishers, academic repositories, proprietary research databases, niche information services — any asset with real value that was previously un-monetizable at small transaction sizes — are now candidates for new revenue streams. This is a category no one is actively building into yet.

The Real Bottleneck Is That Most People Cannot Think Algorithmically — and This Won't Change Quickly

Casado makes a point that is profound in its implications but gets steamrolled by the conversation moving forward: the majority of knowledge workers cannot produce a flowchart of their own job. If agent utilization requires algorithmic thinking, then the actual diffusion rate is constrained not by AI capability but by human cognitive structure.

"If you were to go into any person and ask them to create a flow chart for a particular thing that they have to go do, they would probably fail at producing that flow chart... Within any organization, say, doing a marketing plan and there's 50 marketing people working on a giant product line — one person probably understands and could document the flow chart." — Martin Casado 00:03:04

The overlooked implication: the companies that will win the agent era earliest are not those with the best AI tools, but those with the highest concentration of systems thinkers who can actually direct agents. This reframes talent strategy entirely — the scarcest resource is not AI access, it's the human capacity to decompose work into agent-executable logic. Companies and investors should be looking for this cognitive profile as the new "10x engineer" of the agent era.