Misryoum reports on Aaron Levie’s take: AI agents may expand software demand across every enterprise function—but only if companies modernize data, context, and where automation should be deterministic.
Aaron Levie’s message at O’Reilly AI Codecon wasn’t a victory lap for software jobs—it was a reality check about how enterprises will have to change.
In that spirit. the central tension Misryoum sees is simple: software engineering demand is climbing. AI roles are surging. and yet the hardest part of the AI transition may not be hiring.. It’s deciding what gets automated. what stays judgment-driven. and how to structure enterprise data so agents can actually use it.
The “engineering demand paradox” is real—if you believe in diffusion
Levie’s framing leans on a principle that resonates beyond policy arguments: if AI agents make a software engineer dramatically more productive. then projects that used to be too expensive become feasible.. That doesn’t just increase work inside tech companies; it spreads demand across the wider economy.. In his “Jevons paradox happening in real time” framing. the result is not fewer builders—it’s more organizations discovering they can finally justify building the systems they always wanted.
For workers, this matters because the job label may stay “engineer,” but the mission changes.. Instead of being confined to IT. engineers (and the agent workflows they help create) can move into marketing operations. legal workflows. accounting close processes. or internal audit.. Misryoum has also watched how badly enterprises still run on manual handoffs and fragmented systems—so “more software” often means “less overhead” and “fewer steps between intent and action. ” not just more headcount.
Context beats connectivity—agents don’t win on wiring alone
Levie’s warning is blunt: if information is scattered across dozens of systems and isn’t structured for agentic use, the outcome becomes less like automation and more like probability. Misryoum interprets that as a practical risk for enterprises: “connected” isn’t the same as “actionable.”
This is where infrastructure modernization enters.. It’s not a flashy headline. but it’s the work that determines whether agents behave reliably or stumble through workflows.. In the enterprise world—thousands of employees. hundreds of legacy applications. layers of proprietary data—getting the context right tends to be measured in months. not weeks.
There’s also a human dimension here.. When agents fail to find context, teams often respond by adding more manual review.. That can create a new kind of friction: employees end up double-checking both the agent’s output and the underlying assumptions.. The cost isn’t only technical; it’s operational and cultural.
Deterministic code vs probabilistic AI: the “trillion-dollar” boundary
Some processes must be stable.. Loan processing and other compliance-heavy operations typically require repeatability: the same inputs should lead to the same outputs, every time.. But employee support questions. early-stage classification. or exploratory analysis may tolerate uncertainty—especially if the user interface supports review and clarification.
Levie described this as a “trillion-dollar question,” and Misryoum agrees because the decision touches cost, risk, user experience, and scalability.. Even when an LLM can do something, canned code might be cheaper and more reliable.. On the other hand, liquid user experiences can be worth the extra inference cost when flexibility is the point.
What’s often missed in public AI rhetoric is that better automation requires more technical sophistication, not less.. AI may simplify some tasks. but it raises the skill bar for system design—how tools are orchestrated. how data is prepared. and how failures are handled.. In other words. the “two computers” metaphor isn’t academic; it’s a blueprint for how modern products will be built.
Startups can win—especially where incumbents don’t own messy workflows
That’s why services domains are becoming fertile ground for AI-native firms.. Levie suggested that in many of these areas, the main “incumbent” is professional services rather than large enterprise software vendors.. If a startup can start fresh. design workflows around agents. and deliver output faster with lower costs. incumbents may find it difficult to respond quickly—especially if their own organizations are built around older process constraints.
But there’s a risk on the enterprise side too.. Misryoum sees it frequently: teams try to “stuff AI into” existing workflows instead of redesigning the workflow for an AI-first reality.. Organizations are attached to roles and org charts, and job functions can become identity.. Without honest reengineering. AI adoption can turn into a costly layer on top of systems that were never structured for agentic context.
The most valuable “asset” in the enterprise might be invisible context
In Levie’s model, agents can be thought of as new expert employees who arrive with zero context. They’re capable, but they need briefing—and the briefing must be precise. Too much information can confuse, while too little forces the agent to roll the dice.
Practical solutions like SKILLS.md and AGENTS.md attempt to codify the smallest set of context needed for a specific process.. Misryoum’s takeaway is that this approach works best when the organization can standardize workflows and define “just enough” structure.. For the rest of knowledge work—where “99% of processes” lack an AGENTS.md equivalent—the challenge becomes reengineering.. The information is everywhere, but the interface between that information and the agent is not.
Levie’s Box pivot summary captured the direction of travel: move from “content” to “context.” For enterprises. that means treating contracts. research materials. and financial documents not as static files. but as inputs that must be prepared for agent decision-making.. Misryoum expects this to be the real competitive line over the next decade: not who has the best model. but who can turn their business reality into the structured. timely context agents need to act.
Anthropic launches Claude Design—can it outdo Figma?
Old iPad or Android tablet as a smart home panel: the setup guide
OpenAI exits: Kevin Weil and Bill Peebles leave as “side quests” end