Karpathy vs. McKinsey: The Truth About AI Agents (Software 3.0)

AI agents are colliding with enterprise reality, and the sharpest contrast this week came from Andre Karpathy’s “Software 3.0” framing versus McKinsey’s “agentic mesh” pitch. Karpathy’s core claim is that large language models should be treated less like deterministic programs and more like “stochastic simulations of people”—useful, human-feeling, and powerful, but inherently jagged in execution. That framing matters because it forces a design shift: software must assume humans will validate outputs, and systems must be built to keep that human-in-the-loop workflow sustainable.

Karpathy’s talk to entrepreneurs at Y Combinator Startup School argues that the next coding language is effectively English, not because code will disappear, but because LLMs behave like probabilistic “people spirits” rather than reliable machines. He compares LLMs to utilities—metered by tokens like electricity—and to operating systems, where different model ecosystems create user preference “wars” similar to Windows versus Mac. The practical takeaway is about how to build for agents that can’t yet be trusted to execute high-stakes tasks end-to-end. Instead of betting on full autonomy, teams should design validation loops first, making checking as easy as possible. He also recommends constraining generation—“putting the LLM on a short leash”—so evaluators aren’t overwhelmed by hundreds of candidate outputs when humans can only verify a small subset.

The transcript also flags where Karpathy’s view may be incomplete. English may become a dominant interface, but complex systems still require strong technical engineering, especially as traditional software interacts with agentic, AI-augmented components. The talk acknowledges the limits of “vibe coding” (popularized by Karpathy’s earlier work): it works well in local environments, but breaks down across deployment pipelines, CI/CD, and integrations. Karpathy’s broader vision is an “augmented Iron Man suit” for human reach and control—agents expand what people can do, but data systems, control systems, and validation mechanisms must be engineered to match how probabilistic models behave.

That builder-first emphasis is contrasted with McKinsey’s CEO-facing messaging. While the boardroom-friendly themes—workflow thinking, not just task automation—are directionally correct, the transcript criticizes “agentic mesh” as vague “word salad” lacking empirical grounding. The concern is practical: McKinsey’s implied promise that agents can be plugged in like USB ports, swapping models such as “Mistral small” or “GPT-3.5 Turbo” with minimal rework, clashes with how enterprise systems actually ship. The transcript points to a key mismatch between simplified narratives and engineering constraints, including the underwhelming performance of edge computing for models—an approach that has not delivered the expected sustained gains.

The closing message is a call for cultural change: organizations need to tell the truth about AI system complexity and adopt a crawl-walk-run rollout strategy rather than starting with full automation of core business lines. The payoff can be real—agentic systems can unlock major value—but only if teams design for human validation, constrain generation, and build systems that don’t pretend integration is plug-and-play.