A compiler judges code in milliseconds; even the fastest engineer review takes minutes — a million-fold gap. We focus on this gap because it precisely explains why AI Agents (AI that autonomously executes tasks) can write and modify code today, yet stall in healthcare, law, and enterprise management.
What This Is
The original author points to two reasons.
First, can knowledge be "compressed"? In code, x + 1 is always x + 1 — no contextual dependency. But medicine's "suspected" and "cannot rule out," law's "duty of reasonable care," and management's "right moment" — ambiguity itself is the most precise form of knowledge. Force it into deterministic rules and you lose the information itself.
Second, is there a fast, objective "referee"? Coding has compilers and unit tests — a try → feedback → adjust reinforcement learning loop (training that lets AI improve through trial and error) runs in milliseconds. Other domains? Doctors wait weeks for clinical outcomes; lawyers wait for disputes to arise; managers wait half a year to see results. Break the feedback chain and Agents remain stuck at "imitating human output," never crossing the threshold of "learning from mistakes."
GitHub is special because commits, PR reviews, and unit tests naturally form a complete causal chain — and automated model judgments on a 100,000-star project are reliably accurate. This structure simply does not exist in medical records, court rulings, or ERP systems.
Industry View
This is not industry consensus. Optimists argue: Claude's Computer Use and OpenAI's Operator are standardizing "tool calling," and Cursor's $10 billion valuation is itself the market's bet on a vertical Agent explosion.
But skepticism is equally sharp. The original author's take: replicating the coding path requires a "domain-level GitHub" — which healthcare, law, and enterprise management cannot build in the short term. Over the next five years, Agents will still land primarily in domains that have "a compiler as referee."
An even cooler counterargument: the original author underestimates the human feedback path. Anthropic and OpenAI use RLHF (Reinforcement Learning from Human Feedback) for alignment — the problem is cost. A senior doctor takes half an hour to judge whether a prescription is correct, costing hundreds of dollars at market rates. To train a usable vertical Agent, the required feedback data volume is astronomical.
Impact on Regular People
For enterprise management / IT: when evaluating AI projects, ask first — does this scenario have a "millisecond-level objective feedback mechanism"? If yes, it's worth trying. If not, the demo will likely look pretty but won't stick.
For individual careers: rather than chasing an "all-purpose Agent," identify the "verifiable steps" in your own work. Code, report reconciliation, contract clause matching — AI can boost these immediately. Judgments that require balancing ambiguity are beyond AI's near-term reach.
For consumer markets: "AI Agent" products in healthcare, law, and education will mostly remain marketing packaging in the short term. Before paying, ask one more question: "Who bears the consequences when the judgment is wrong?" — a better bubble detector than any demo.