August 27 2025
In a recent conversation on the a16z YouTube channel, Aaron Levie (CEO, Box) and Steven Sinofsky (a16z board partner; former Microsoft exec) joined general partners Erik Torenberg and Martin Casado to explore one of the hottest debates in AI: what exactly is an “agent” — and how will agents change the way we work?
The discussion unpacked both the technical foundations and the economic implications of AI agents, comparing today’s systems to “bad interns” while pointing toward a future of specialized, autonomous collaborators.
At its simplest, Sinofsky quipped, an “agent” is like the Linux background task operator — something quietly working behind the scenes. Levie pushed further: the more an AI system can run independently, take action, and feed its outputs back into itself, the more “agentic” it becomes.
Yet today’s agents are not singular AGIs. Instead, they resemble networks of narrow sub-agents, each specializing in discrete tasks. This modularity mitigates context loss and mirrors the division of labor humans have long used to increase efficiency.
The panel repeatedly drew on the history of computing. Just as early PCs shifted workflows from manual processes to digital ones, AI agents may force us to conform our work to how agents operate — not the other way around.
Sinofsky pointed to past platform shifts: the emergence of GUIs (graphical user interfaces), the rise of the internet, and the transformation of productivity software. Each wave began by mimicking old processes (word processors printing on pre-formatted expense sheets) before inverting them entirely (cloud-based systems like Concur automating the workflow itself).
AI, they argued, is poised to follow a similar path — starting clumsily, but eventually reshaping how work itself is organized.
Agents already enhance productivity, but their greatest value often accrues to experts. Skilled engineers, for example, can harness code-writing agents to achieve 10x productivity gains, while novices risk deploying flawed outputs.
This dynamic suggests not the erasure of human expertise, but its amplification. Just as specialization expanded in medicine, software, and construction over the last century, AI may accelerate the creation of new professional roles: “AI productivity managers,” agent orchestrators, and domain-specific supervisors who direct networks of specialized systems.
Looking ahead, Levie and Sinofsky see opportunity not in building monolithic “do-everything” agents, but in developing verticalized, applied AI. Each industry — from law to healthcare to sales — could see its own wave of agent-driven startups, much as APIs once became standalone companies.
At the platform level, the shift is profound: for the first time, application logic is being abdicated to third-party models. Just as operating systems once standardized drivers and browsers standardized interfaces, large language models may redefine the layers on which new enterprises are built.
According to these experts, what emerges is less the fantasy of AGI “replacing jobs” and more a reconfiguration of workflows, roles, and organizational design. Agents may not learn how we work so much as we learn to work differently because of them.
As Sinofsky concluded, every major shift in computing has begun with awkward experiments and ended with entirely new paradigms. AI agents, still in their “bad intern” phase, may ultimately prove to be the next enduring platform — one that transforms both the tools we use and the very structure of work itself.
