# Agentic Engineering Is the New Dev Discipline — Here’s What It Requires > Source: > Published: 2026-08-12 07:12:11+00:00 Andrej Karpathy coined “vibe coding” in February 2025 and killed it almost exactly a year later. At Sequoia Capital’s AI Ascent event in February 2026, he introduced the term replacing it: **agentic engineering**. Not a tool. Not a framework. A discipline. Most developers using AI agents today are still vibe coding — they’ve just added more steps and convinced themselves otherwise. ## What Agentic Engineering Actually Is Karpathy’s definition is precise in a way his original vibe coding post was not: “The professional discipline of coordinating fallible agents while preserving correctness, security, taste, and maintainability.” The critical word is *fallible*. Vibe coding assumed the AI would mostly get it right. Agentic engineering begins from the opposite premise: the agent will make mistakes, and your job is to architect the system so those mistakes are caught before they compound. Five skills define the discipline: **Spec design**— write formal requirements before any prompting begins** Plan supervision**— review the agent’s generated plan before it executes** Diff review**— read every generated diff for architectural correctness, not just functional correctness** Eval loop design**— build feedback loops with verifiable, automated completion signals** Security oversight**— manage permissions, isolation, and trust boundaries explicitly This is not a checklist you run once. It’s a workflow you internalize. The developers getting hired for agentic AI roles aren’t the ones who prompted their way to a demo. They’re the ones who can answer: “What’s your eval loop?” and “Show me your AGENTS.md.” ## The Spec You’re Not Writing Most developers go straight to prompting. [Spec-driven development](https://www.augmentcode.com/guides/what-is-spec-driven-development) says: don’t. A spec is 10 to 30 lines. It defines goal, requirements, constraints, and acceptance criteria. It takes five to ten minutes to write. According to AWS Kiro’s documented customer cases, teams writing specs first are shipping 40-hour features in under eight hours of human time. The mechanism for persisting specs across agent sessions is the standing instruction file — CLAUDE.md, AGENTS.md, and their equivalents. Karpathy noted that frontier models reliably follow roughly 150 to 200 standing instructions before compliance starts to degrade. That’s a feature, not a bug: a lean, precise spec outperforms a sprawling one. “Spec until the prompt is obvious” is the practical reframe. The tooling has caught up. AWS Kiro is built around this workflow. GitHub’s Spec Kit reports teams shipping with roughly ten times fewer “regenerate from scratch” cycles compared to ad-hoc prompting. The spec isn’t extra overhead. It’s the part that makes the agent reliable. ## Eval Loops: The Gap You Probably Have Vibe coding has no systematic verification. You look at the output, it seems right, you ship it. Agentic engineering has a Stop Hook. A Stop Hook intercepts the agent’s exit attempt, checks whether completion criteria are actually met — tests passing, coverage threshold hit, type checks clean — and reinjects the task if they’re not. Claude Code shipped `/goal` in May 2026: set a condition, Claude works autonomously until it’s met. This is not magic. It’s engineering. The gap in current practice is stark. Seventy-one percent of developers say they won’t merge AI-generated code without manual review. But only 38% have automated eval loops. The other 33% are doing manual review at scale — which is not a sustainable workflow, it’s technical debt in the form of human attention. The harder evaluation challenge is trajectory validation. Unlike standard functions, agentic loops can’t be unit-tested against fixed input/output pairs — the agent has autonomy over sequencing. Production deployments break not because the model reasons badly, but because of gaps between tool calls, retrieval, and the audit trail. [TechEmpower’s July 2026 analysis](https://www.techempower.com/blog/2026/07/14/evals-for-agentic-loop-applications/) found that the most reliable agentic systems use a three-pillar eval: final output quality, trajectory analysis, and trust/safety scoring at each step. ## Security Is Not Optional The data is not ambiguous. [Veracode’s 2026 GenAI Code Security Report](https://www.veracode.com/blog/2026-genai-code-security-report-ai-risk/) found that 45% of AI-generated code contains security vulnerabilities. Forty-one percent of AI-generated backend code includes overly broad permission settings. AI-generated code now causes one in five enterprise security breaches. The concrete agentic engineering practice here is worktree isolation. Claude Code has shipped native worktree support since February 2026 — each agent session gets its own git worktree, its own branch, with zero risk of file-state collision between parallel sessions. Between July 20 and July 24, 2026, Claude Code shipped four consecutive releases closing isolation boundary gaps: symlinked working directories, git redirection out of a worktree, leftover worktrees from other projects, and unprompted network egress. Those four releases happened because developers deployed agents without thinking through isolation boundaries first. Security oversight in agentic engineering means minimum permissions, explicit trust boundaries, and isolation by default — not as an afterthought. ## What the Job Market Is Telling You If the discipline argument doesn’t land, the hiring signal should. There were 90,000 agentic AI job postings in 2026, up 280% year-over-year, with an [average salary of $190K](https://jobsbyculture.com/blog/agentic-ai-hiring-boom-2026). Job postings mentioning agentic AI skills have grown 986% since 2023. What hiring managers in 2026 actually want is not “uses AI tools.” Every candidate uses AI tools. They want measurable impact from AI-assisted work, reliable guardrails, reproducible workflows, and documentation fluency — meaning you can show them a CLAUDE.md or AGENTS.md that reflects deliberate choices, not defaults. The spec and the eval loop are what separate the candidate pool. ## What to Do This Week Three concrete steps, in order of impact: **Write a CLAUDE.md or AGENTS.md for your current project.** Ten to thirty lines. Define the project goal, the invariants the agent must respect, and the security boundaries it cannot cross. This alone reduces “agent went sideways” incidents significantly.**Add one Stop Hook.** Pick the most important completion signal in your current workflow — tests passing, a lint check, a type check — and implement a hook that prevents the agent from exiting until that signal is green.**Audit your last five merged AI-generated diffs.** Would you have caught a permission misconfiguration? An overly broad database query? If the answer is probably not, your diff review process needs structure, not speed. Vibe coding was a useful gateway. It lowered the floor and let more people build. But the floor is not where production software lives. [Agentic engineering is the upgrade](https://karpathy.bearblog.dev/sequoia-ascent-2026/) — and based on the hiring data, the teams that internalize it first will have a significant head start on the ones still iterating by feel.