TBPN Sep 21, 2026 27m 16m saved
With Aaron Levie, co-founder and CEO of Box
Box has re-accelerated its growth past its own internal plan, and Aaron Levie's explanation is the opposite of the argument that was current six months ago.
The consensus then was that agents would hollow out application software: once a model can do the work, all you need underneath is a database. Levie's position is that the database is precisely the thing that gets harder, because it now has to arbitrate between thousands of people and tens of thousands of agents at once.
"And actually, the more agents that you have running on these systems, in fact, the more important the backbone becomes."
Levie co-founded Box and has run it since, and the company sells the system enterprises keep their contracts, financial documents, research materials and marketing assets in — which is exactly the material an agent has to reach to do anything useful inside a large organization.
The full interview is covered here so you can skip it. 27 minutes of audio, 11 minutes of reading.
Here are the 8 arguments that matter.
Key Takeaways
Box was among the first to call the software panic over, and Levie says you could not have called it much earlier without missing the bottom
Agents raise the coordination problem, they do not remove it — a store now answers to thousands of users and tens of thousands of agents
Box has "re-accelerated growth far past our internal plans", which he attributes to enterprises needing an authoritative place for unstructured data
Nothing is being run the way it was 18 months ago, and some categories will still lose while others gain more than they lost
Search had to be rebuilt for a reader that is not human: embeddings, agent re-ranking, parallel queries
The Hugging Face incident changed how he thinks about internal security, not just production systems
He was wrong about consumer AI and says so — the breakthrough was packaging, and it was visible in OpenClaw all along
Authentication is the missing piece, not payments, which he says Stripe and others have largely handled
AI becomes the dominant 2028 election subject, and the industry has not decided what it wants
1. Calling the Bottom Early
The hosts opened on the software selloff and the newsletter Box published during it declaring the panic finished. Levie said the timing was deliberate in both directions: they moved fast, but you could not have called it before the market had fully capitulated.
What the capitulation established, in his telling, is that the systems of record were not the weak link.
More agents on a system make the system matter more
And actually, the more agents that you have running on these systems, in fact, the more important the backbone becomes.
Aaron Levie
The bear case he was answering is that agents only need a database underneath them.
The database is the hard part, not the simple part
But the problem is that database has to be coordinated against possibly thousands of end users and then possibly tens of thousands of agents.
Aaron Levie
Access controls, business logic, rendering the experience and security all get harder rather than easier under that load, he said, and the Hugging Face incident at OpenAI is the demonstration.
His own weekend supplied the consumer version of the argument. Moving AI-generated videos from a PC to a phone, he never considered having the model build him a storage system.
Nobody wants a one-shot clone of infrastructure that already works
I don't want to waste a prompt. I don't want to waste a second doing that. Let's use what I already use.
Aaron Levie
I don't need you to one shot some shoddy clone. It's a waste of time. It's a waste of money. It's a waste of energy.
Aaron Levie
2. Not Business as Usual
A host asked how much of the recovery is investors deciding the threat was imaginary versus deciding the opportunity outweighs it. Levie rejected the first framing outright.
Every software company is being run differently now
I don't know of a single company still being run the way that it was run maybe, 18 months ago.
Aaron Levie
Some parts of the market will capture less value than before, he said, and some will capture a great deal more. What he will not accept is the idea that nothing changed.
3. Where Agents Create Value
Asked which businesses are on the winning side, Levie named the layers and the reason for each.
Infrastructure providers are the clearest case: agents need sandboxes, compute, network and gateways. Data platforms are the second, because an agent triaging large datasets needs them pulled into one system.
Box's own growth is the third example
For us, we have re-accelerated growth far past our internal plans because it turns out that enterprises need core systems to be able to manage their unstructured data.
Aaron Levie
The personal-productivity intuition — just keep the file in a cloud folder, or on the device — does not survive contact with a company of ten thousand people.
At enterprise scale the question is which copy is the real one
How do you ensure that those agents continue to go after the actual canonical records, the actual documents that are the real sources of truth and that are the authoritative versions of that data? That means you need platforms.
Aaron Levie
His overall verdict on what agents have done to software
It has unequivocally been a net positive for software, even with some of the spaces being under pressure
Aaron Levie
The reason is that the platforms now have a second job: guardrails, traffic control for what the agents are doing, and the ability to run agents natively inside the system of record.
4. Search Built for Agents
A host asked what the biggest lever is — latency, because a hanging agent is worse than a person waiting for a search, or sheer volume, because everyone and everything is producing more files. Levie said both, and added a third.
Content creation is exploding on its own, he said, and agent-produced material has to be stored, governed and managed exactly as if a person had made it.
The layer above that is retrieval, which Box has rebuilt. Documents are captured as embeddings, agents re-rank the results, and multiple versions of a query run in parallel.
Retrieval is faster and more accurate because it is parallel, not just smarter
Your chance of finding your document not only faster but just generally more accurately is way higher today than it was a couple years ago because we not only have all these other signals, but we can parallelize the process much better.
Aaron Levie
An agent arriving from outside and asking for a specific record, he said, now gets it.
5. Securing Data From Agents
The third area is security, and the reference point he used was the Hugging Face incident — a production system under threat, with implications he thinks run straight into the enterprise.
The internal environment is the next version of that problem
If you have all of your critical contracts or financial documents in an environment that agents can eventually get their way into, enterprises are going to need to care deeply about how do you protect that information.
Aaron Levie
Box is redesigning data protection for an environment with agent swarms deployed against it, working from the assumption of a hundred times more agents than people and asking how you would store, access, search and secure anything under those conditions.
A host's contribution was linguistic: the industry should stop saying sandbox, on the grounds that no parent has ever watched a five-year-old stay in one. Levie agreed the term has run its course.
6. What He Got Wrong
On personal agents, Levie opened with a concession rather than a view.
He did not see the consumer breakthrough coming
I want to first apologize because I honestly did not have the imagination on the consumer front.
Aaron Levie
He had been skeptical about what a consumer breakthrough would even look like, and says the answer was in front of everyone.
The capability was already visible; the packaging was the innovation
And it's funny because it was obviously staring at us with OpenClaw.
Aaron Levie
What Muse and instinct did was the form factor. What made it work underneath is the progress on computer use and browser use, which he called amazing.
His architectural wish list has one item on it.
Payments are done; authentication is not
Payments are largely solved because of obviously the work that Stripe and others have done.
Aaron Levie
What he wants is a one-click, secure way into any product, with a trusted handshake between the agent and the tools — and he says authentication is the piece still to be built as agent use ramps.
7. More Demand, Not Less
Levie's economic read on personal agents is expansionary rather than disintermediating.
Handing over the whole task is the consumer breakthrough
The ability to now hand off the whole problem to an agent as opposed to just using it nearly as a research assistant, I think is going to be a breakthrough for consumer use cases.
Aaron Levie
His own example is local services. The florist he uses makes excellent arrangements and has an ordering process he described as painful, and that friction is the reason he buys less than he otherwise would.
Lowering the friction raises his consumption, not just his convenience
I might now be a consumer of five times the amount of local service providers as a result of these types of systems.
Aaron Levie
He conceded the counter-framing offered by a host — that this is cutting out the middleman and installing a more powerful one — and answered that it may well be middlemen all the way down, with the graph of the local economy as the asset.
On the objection that consumers do not value productivity, Levie said he had held the same view and now thinks it is the wrong axis. The pitch is not saving time; it is doing more. He has three children, and the list of camps, museums, rentals and sign-ups only grows.
The output is less stress and more activity, not fewer hours
So I'm a big fan of the personal agents. I think it's just full net growth in the consumer economy.
Aaron Levie
A host put the bear case: that this turns out to be a prosumer moment that never crosses the chasm, and those companies pivot to enterprise deployments. Levie's answer was that the required behavior change is unusually small.
No new hardware, no new habit — it is the phone people already hold
The form factor itself is just a mobile app on your phone.
Aaron Levie
If the technology reaches something near reliable execution, he said, this is not the kind of shift that needs society to adopt headsets.
8. The Safety Language Trap
Asked for his probability of catastrophe from AI, Levie put himself in an unusual position: low estimate, high appreciation of the safety work.
The attention on safety is the reason his estimate is low
And I think that because of the amount of attention on the safety, I think that's sort of what causes me to have a low P. Doom, ironically.
Aaron Levie
He sees events that could be catastrophic, he said, and thinks the ecosystem has a strong incentive to go after them, which lowers the probability. A host made the same point from the other side: most published estimates are conditional on doing nothing, so the number after countermeasures is the one that matters.
Levie's worry is downstream of the language rather than the risk.
Safety vocabulary gets picked up by people with different goals
I think the one thread that we still have to figure out, though, is the language will be co-opted by parties that don't have the same kind of either level of either view or goals or incentives.
Aaron Levie
The narrow line he described runs between elevating the risk enough to fund alignment work and elevating it enough that voters simply switch the thing off. If the industry tells people the danger is high and the current consumer benefit is an agent that shops for them, he expects a lot of people to take the off switch.
He also does not think the industry knows what it wants.
One week an open-source letter, the next week a pacing letter
You have this sort of roller coaster of, one day everybody signs the open-source and then the next day everybody signs the pacing letter.
Aaron Levie
Some people can reconcile the two, he said, but they pull in opposite directions.
Where he expects this to land politically
I think it's going to become the dominant 28 conversation from a presidential election standpoint, like it has to be because it's going to deal with jobs and GDP and safety and the next sort of major risk that people start to think about.
Aaron Levie
So I expect this to be quite messy for the coming years.
Aaron Levie
Bonus Insights
He almost put a deposit on the Tesla Roadster this time
Levie said he skipped it nine years ago and was tempted now, having gone down a twenty-minute video rabbit hole of people whiting out the teaser footage to find hidden text.
His skepticism about the flying-car specification is practical, not technical
No, I guess the question is like, practically, what does that do? If you literally can't fly it.
Aaron Levie
A host noted the original Roadster specification — zero to sixty in two seconds, a $250,000 price, more than 600 miles of range — and that the range figure is the one no electric vehicle has met.
On Elon Musk's product choices
We may have reached the point, though, where Elon is building things for his personal life that exceed our normal zoning laws of mere mortals.
Aaron Levie
An agent booked his family's day out
He described booking the Japanese tea gardens in San Francisco through an agent and adding the outing to the day simply because it had become easy enough to execute.
A host's framing of what this feels like
One host relayed a friend's description of working with agents as a return rather than a novelty: dictating, delegating and taking memos the way a Gilded Age executive did, now available to everyone instead of a fraction of a percent.
Levie's bottom line is that agents have not devalued enterprise software, they have raised the bar for it — the systems that hold the authoritative version of a company's data become the control point precisely because everything else is being automated around them.
Products, Companies & Tools Mentioned
Box (Levie's company. Growth re-accelerated past internal plans on demand for a governed home for unstructured data; rebuilt search with embeddings, agent re-ranking and parallel queries)
Cloudflare (His example of an infrastructure provider doing well because agents need sandboxes, compute, network and gateways)
Databricks and Snowflake (Named as beneficiaries because agents have to triage large datasets in one place)
Hugging Face and OpenAI (The security incident he treats as the template for what an enterprise agent environment has to defend against)
Stripe (Credited with largely solving agent payments, against authentication, which he says is still missing)
Meta and Muse (One of the two consumer agents whose packaging he says he failed to anticipate)
Yelp, Thumbtack and TaskRabbit (His shorthand for the local-services graph he expects agents to expand rather than disintermediate)
Tesla (The Roadster deposit he nearly placed, and the 600-mile range figure he says no electric vehicle has hit)
Listen to the full episode
🔴 YouTube | 🔗 Episode page
Watch the full episode:
More Aaron Levie interviews and podcast appearances
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: