I sat down with Aaron Levie on the O’Reilly AI Codecon two weeks in the past. Aaron cofounded Field in 2005, and 20 years later, his firm manages content material for about two-thirds of the Fortune 500. Aaron is among the few CEOs of an incumbent enterprise software program firm considering deeply in public about what AI means for all the enterprise stack. There are lots of people who’re constructing corporations from the bottom up with AI, others who’re dragging their ft adapting current enterprises to it, after which there’s Aaron. He sits in a form of Goldilocks zone, enthusiastic however not uncritical, partaking within the arduous work of adapting AI to the enterprise and the enterprise to AI.
The engineering demand paradox
I began out by asking about one thing from Lenny’s E-newsletter that Aaron had retweeted. Regardless of all of the doom rhetoric, TrueUp knowledge reveals software program engineering job postings are at a three-year excessive. Product supervisor jobs are approach up. AI jobs as a complete are approach up.
The precise knowledge could also be extra equivocal than the TrueUp report suggests. The trustworthy learn of the literature as of spring 2026 (Brynjolfsson et al., Humlum and Vestergaard, BLS Software program Builders, BLS Laptop Programmers) is that one thing actual is occurring to entry-level software program work, that it’s taking place sooner than most earlier expertise transitions, that it has completely different results relying on which job code you have a look at, and that it isn’t but clear whether or not the web impact on whole software program employment might be detrimental, impartial, or finally constructive. Nonetheless, the TrueUp report was a set off for the dialogue that adopted.
Aaron famous that engineers have traditionally been concentrated at tech corporations as a result of the price of a software program undertaking was too excessive to justify wherever else. But when brokers make an engineer two to 10 occasions extra productive, all of the software program tasks that have been by no means economically viable immediately turn into viable. Demand doesn’t shrink. It diffuses throughout all the financial system. In his tweet, he referred to as it “Jevons paradox taking place in actual time.” In our dialog, he mentioned:
“What’s going to occur is all the world goes to be taking a look at all of the potential software program that they construct. They usually’re going to begin to say, Oh, I can lastly justify going out and doing the sort of undertaking the place I couldn’t earlier than.”
Engineers empowered by AI brokers received’t simply construct software program for IT groups. The entire addressable position of the engineer expands from the expertise division to each operate within the enterprise. They’ll be wiring up automation for advertising, authorized, accounting, and each different company operate.
He’s completely proper. Go searching in any respect the crappy workflows, the crappy processes, the unimaginable overhead of issues that must be easy. You suppose corporations ought to lay off their builders to cut back prices when there’s a lot shitty software program on the market? Actually? There’s a lot that must be improved. He had a fantastic line: “Silicon Valley is spooked by its personal expertise.”
Over to me: The rhetoric from the labs about job destruction is actively counterproductive. I used to be speaking just lately with somebody in healthcare who described a hospital system attempting to fill an enormous gap from lowered Medicare funding. They see AI as a technique to achieve effectivity of their again workplace to allow them to liberate extra sources for affected person care. And naturally the union is combating it as a result of they’ve been informed AI is a monster that’s going to take their jobs. If you happen to inform a unique story, one about making the system higher and serving extra individuals extra affordably, that’s one thing individuals can get behind. We now have to alter the narrative.
Context, not connectivity, is the actual drawback
I additionally requested Aaron whether or not protocols like MCP are making context moveable sufficient to erode aggressive moats. He agreed that the business has broadly converged on openness and interoperability (with some toll cubicles to work by means of). However getting your techniques to speak to one another doesn’t resolve the more durable drawback of getting your knowledge structured in order that brokers can really discover the fitting data on the proper second.
“If it’s in 50 completely different techniques and it’s not organized in a approach that brokers can readily reap the benefits of, what you’re going to be is on the mercy of how effectively that agent finds precisely the context that it must do its work. And also you’re form of simply rolling the cube each time you do a workflow.”
He predicts a decade of infrastructure modernization forward, which sounds about proper. At O’Reilly, I preserve working into this myself. I’ll see a job that’s good for an agent and shortly uncover that the info I want is scattered throughout 4 techniques and I’ve to leap by means of hoops to determine who is aware of the place the info is and easy methods to get entry. A buddy working a big (however comparatively new) enterprise that’s turbocharging productiveness and repair supply with brokers informed me just lately {that a} huge a part of his workforce’s success was doable as a result of that they had spent numerous time getting their knowledge infrastructure so as from the beginning.
IMO, numerous the tales you hear about OpenClaw and different harbingers of the agent future could be deceptive in an enterprise context. They’re doing greenfield setups, largely working client apps with well-defined interfaces, and even then, it takes weeks to arrange correctly. Now think about agentic frameworks for corporations with 1000’s of staff, a whole lot of legacy apps, and deep wells of proprietary knowledge. A decade of infrastructure modernization is beneficiant. With out assist, many enterprises can have problem making the transition.
Engineering the trade-offs
I introduced up Phillip Carter’s “two computer systems” framing, that we’re now programming a deterministic laptop and a probabilistic laptop on the similar time. Abilities are a bridge, as a result of they’ve each context for the LLM which might work probabilistically and instruments which might be constructed with deterministic code. Each techniques coexist and work in parallel.
Aaron referred to as the boundary between the 2 computer systems “the trillion-dollar query.” When does a course of cross the brink the place it needs to be locked into repeatable, deterministic code? When ought to it keep adaptive? Mortgage processing must work the identical approach each time. Worker HR queries could be probabilistic. And the irony, as Aaron identified, is that making these trade-offs appropriately requires deep technical understanding. AI makes the sector extra technical, not much less.
I added that generally this judgment is a person expertise query, generally a price query. You are able to do one thing with an LLM, but it surely is perhaps loads cheaper with canned code. At different occasions, despite the fact that the LLM prices extra, the pliability of a liquid person interface is much better.
That is additionally a locus of creativity. What you carry out of AI is what you carry to it. Steve Jobs wasn’t a coder, however he knew easy methods to get essentially the most out of coders. He would have gone nuts with AI brokers, as a result of he was the essence of style and judgment and setting the bar.
The place startups win
I requested Aaron concerning the dangers to current enterprises from greenfield AI startups that may simply transfer sooner, reinventing what the incumbents do with an AI native answer, with out all the bags. He replied:
“If there’s already a considerable quantity of the info for that individual workflow in an current system, and the incumbent is agile sufficient and responsive sufficient, then they’re in a great place to construct both the options or to monetize that set of labor that’s going to be completed….What brokers are actually good at is automating the unstructured areas of labor, the messy, collaborative human-based components of labor, the tax course of, the authorized overview course of, the audit and danger evaluation strategy of your whole contracts and unstructured knowledge. And so in these areas, there’s no incumbent. The one incumbent is probably going skilled companies companies. In order that’s the place I might favor startups.”
Software program startups like Harvey are already taking companies domains and constructing brokers for them. However it’s not simply software program startups. Aaron additionally sees numerous alternative for AI-native regulation companies, accounting companies, and advert companies that may throw away legacy workflow, begin from scratch, and ship two to 5 occasions the output at decrease value can have an enormous benefit.
I did push again with some extent I feel is underappreciated: Current enterprises face an actual danger that the group will attempt to stuff AI into current workflows moderately than asking what the AI-native workflow can be. Individuals are connected to their jobs, their roles, the org chart. We now have to wrestle with that actually if we’re going to really reinvent what we do.
People get context without spending a dime
One in all Aaron’s factors about brokers is that people carry an unlimited quantity of ambient context that brokers lack. You recognize what constructing you’re in and who else works there and what they do. You recognize the assembly that simply occurred the place a workforce modified course on a method that hasn’t been written down but. You’ve 20 years of amassed area data. All of that’s free context that we’ve by no means needed to formalize. As he put it, “We’ve by no means constructed our enterprise processes in a mannequin the place we assume that there’s a brand new person in that workflow that appeared one second in the past and in beneath 5 seconds, they should get the entire data doable to try this job.”
He urged that a technique to think about brokers is as new staff who’re consultants however arrive with zero context and should be absolutely briefed. And the context must be exact, not simply complete. Give an agent an excessive amount of context and it will get confused. Give it too little and it rolls the cube. SKILLS.md and AGENTS.md information are makes an attempt to supply precisely the surgical context an agent wants for a particular course of.
However 99% of data work doesn’t have an AGENTS.md file, he famous. The information is in every single place. The context is in every single place. So in an current enterprise, you need to reengineer workflows from the bottom as much as ship the fitting data to brokers on the proper second.
Aaron summed up Field’s strategic pivot in a single sentence: swap the phrase “content material” for “context” and the remainder of the technique stays the identical. Enterprise context lives in contracts, analysis supplies, monetary paperwork. That’s all enterprise content material but it surely isn’t at all times simply obtainable as context. The evolution is making brokers first-class residents alongside individuals as customers of that content material. This very a lot maps to what we’re fascinated by at O’Reilly too.
