Skip to content
a16za16z

Aaron Levie and Steven Sinofsky on the AI-Worker Future

What exactly is an AI agent, and how will agents change the way we work? In this episode, a16z general partners Erik Torenberg and Martin Casado sit down with Aaron Levie (CEO, Box) and Steven Sinofsky (a16z board partner; former Microsoft exec) to unpack one of the hottest debates in AI right now. They cover: - Competing definitions of an “agent,” from background tasks to autonomous interns - Why today’s agents look less like a single AGI and more like networks of specialized sub-agents - The technical challenges of long-running, self-improving systems - How agent-driven workflows could reshape coding, productivity, and enterprise software - What history — from the early PC era to the rise of the internet — tells us about platform shifts like this one The conversation moves from deep technical questions to big-picture implications for founders, enterprises, and the future of work. Timecodes: 0:00 Introduction: The Evolution of AI Agents 0:36 Defining Agency and Autonomy 1:39 Long-Running Agents and Feedback Loops 4:27 Specialization and Task Division in AI 6:04 Anthropomorphizing AI and Economic Impact 9:10 Predictions, Progress, and Platform Shifts 11:31 Recursive Self-Improvement and Technical Challenges 13: 13 Hallucinations, Verification, and Expert Productivity 16:16 The Role of Experts and Tool Adoption 22:14 Changing Workflows: Agents Reshaping Work Patterns 45:55 Division of Labor, Specialization, and New Roles 48:47 Verticalization, Applied AI, and the Future of Agents 54:44 Platform Competition and the Application Layer Resources: Find Aaron on X: https://x.com/levie Find Martin on X: https://x.com/martin_casado Find Steven on X: https://x.com/stevesi Stay Updated: Let us know what you think: https://ratethispodcast.com/a16z Find a16z on Twitter: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Subscribe on your favorite podcast app: https://a16z.simplecast.com/ Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details, please see a16z.com/disclosures.

Erik TorenberghostMartin CasadohostSteven Sinofskyguest
Aug 24, 202556mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

AI agents will reshape work via specialization and new platforms

  1. The speakers define AI agents less as chat interfaces and more as long-running background tasks that execute work autonomously with periodic human check-ins.
  2. They argue that “agency” increases when systems can feed outputs back as inputs via feedback loops, but doing so safely and reliably is technically hard due to distribution shift and control/containment challenges.
  3. Instead of a monolithic AGI, they expect ecosystems of specialized agents coordinated via orchestration, mirroring Unix-style tool modularity and reducing context-rot failures.
  4. Enterprise adoption is becoming more pragmatic: hallucinations are improving, but success depends on verification workflows, expert users, and better prompting/formalization of intent.
  5. They predict AI will drive workflow redesign, more parallelization, and increased specialization—creating many vertical “applied AI/agent” companies despite fears that model providers will subsume the app layer.

IDEAS WORTH REMEMBERING

5 ideas

An “agent” is best thought of as a background worker, not a chatbot.

They frame agentic systems as long-running processes that operate with minimal interaction—more like Linux tasks running “in the background” than a back-and-forth conversation.

True agency requires controlled feedback loops, which remains a hard technical problem.

Casado emphasizes that feeding an agent’s output back into itself risks drifting out of distribution, and that analyzing convergence/divergence resembles difficult nonlinear control theory rather than a simple “arrow back into the box.”

The near-term winning pattern is many specialized agents, not one general one.

To avoid confusion and compounding errors, teams are decomposing work into narrower tasks and orchestrating specialists—an “anti-AGI” but highly effective approach enabled by strong base models.

Context windows get bigger, yet teams still split work because context quality degrades.

They describe “context rot”: as you stuff more into context, answers become lossier, motivating architectures like “one agent per microservice” with dedicated READMEs and ownership boundaries.

AI boosts experts first; verification remains essential and changes who benefits most.

Enterprises increasingly accept probabilistic outputs, adopting review/verify workflows; experts can exploit the tool like a “slot machine” for large productivity gains, while novices risk shipping plausible-but-wrong work.

WORDS WORTH SAVING

5 quotes

And agentification is just hiring a lot of these really bad interns.

— Steven Sinofsky

It, it really gets to the heart of what it means to use a tool. Like, you know, you put me in front of, like, a 12-inch chop saw and say... like, "Go fix the fence," really, really bad idea.

— Steven Sinofsky

AGI just does basically infinite work for every kind of fear we have and maybe every hope that we have.

— Martin Casado

Because it's exponential, you can't predict it, and it's just folly to sit around and try to predict.

— Steven Sinofsky

Office is basically a format debugger.

— Steven Sinofsky

Defining agents: background tasks, autonomy, long-running executionFeedback loops, self-consumption, and nonlinear control challengesContext rot and task partitioning via sub-agentsPrompting, jargon, and formal languages as efficiency mechanismsHuman-in-the-loop verification and changing enterprise cultureWorkflow redesign: parallelization and agent-managed workVerticalized applied AI, platform shifts, and model-vs-app-layer competition

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome