Image by Daniela Paola
Last week, Andrej Karpathy did something that matters more than any model release: he named the discipline. "Agentic engineering" — the practice of designing constraints, feedback loops, and security boundaries that let AI write reliable code — is now a recognized field with a formal definition. When Karpathy names something, it tends to stick.
This isn't just semantics. A named discipline attracts practitioners, builds curricula, and creates markets. And the market is already here.
The Numbers Tell the Story
Claude Code hit a $2.5 billion annualized revenue run-rate as of March 2026, more than doubling since January. A terminal-based coding agent — not a flashy consumer product — is now generating more revenue than most standalone SaaS companies ever will. Meanwhile, 95% of engineers report using AI tools weekly, with Claude Code overtaking GitHub Copilot as the number one tool in the category.
The adoption debate is over. With 55% of engineers already working with autonomous agents — not just copilots — the market has flipped from "should we use AI?" to "who helps us use it well?"
From Patterns to Practice
Simon Willison's agentic engineering patterns guide dropped at exactly the right time. While Karpathy defined the discipline, Willison documented the craft — test-driven development for agent workflows, structured output validation, human-in-the-loop checkpoints. These aren't theoretical. They're production patterns that working engineers need today.
Hexaware went further with RapidX, a spec-driven SDLC framework that makes agentic engineering the default workflow, not an optional add-on. When enterprise services firms are productizing this, you know the practice has crossed from early adopter territory into mainstream engineering.
What This Means
Karpathy's framing highlights something critical: agentic engineering doesn't reduce the need for engineering skill — it redirects it. The core tenets he identifies — security-first design, anti-slop guardrails, structured oversight — are fundamentally engineering problems. The gap between "using AI tools" and "using them with engineering discipline" is where serious practitioners differentiate themselves.
The 75% of smaller companies that prefer Claude Code over enterprise alternatives tells you something about where the energy is. This isn't a top-down enterprise mandate. It's a bottom-up engineering movement that now has a name, a $2.5 billion tooling market, and a growing body of practice.
The discipline exists. The question is whether you're practicing it or just using the tools.
Happy thinking, Jason