SecondBrain
Ask the Brain
Index/Conceptupdated Sun Jun 21 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Agentic Engineering

agentic-engineeringkarpathyengineering-disciplineharness-engineering

Agentic Engineering

Andrej Karpathy's term for the engineering discipline emerging on top of Vibe Coding. While vibe coding raises the floor (anyone can build), agentic engineering raises the ceiling — preserving the professional software quality bar while going much faster.

Karpathy's definition

"You're not allowed to introduce vulnerabilities due to vibe coding. You're still responsible for your software just as before, but can you go faster? And spoiler is you can — but how do you do that properly? That's the realm of agentic engineering."

The core skills

Per Karpathy, what separates an "AI-native" engineer from a "mediocre" one is:

  • Get the most out of the tools available — invest in your setup (CLAUDE.md, prompts, custom slash commands)
  • Drive bigger projects — don't just chunk-ask; design specs, docs, architecture
  • Hold taste, judgment, design — the agent fills in the blanks; you own user-ID design, security boundaries, abstractions
  • Test with adversarial agents — the proposed hiring exercise is a Twitter clone built by you, then attacked by 10 codecs from another applicant

The 10× claim

Karpathy: "People used to talk about the 10× engineer. I think this is magnified a lot more — 10× is not the speedup. People who are very good at this peak a lot more than 10×."

Triangulates with:

The taste bottleneck

"You're in charge of the taste, the engineering, the design, that it makes sense, and that you're asking for the right things… and engineers are doing the fill in the blanks. And that's currently kind of like where we are."

The agent has near-perfect API recall (so you don't need to remember keepdim vs keep_dims), but you must understand the concepts underneath (storage vs view in tensors) to direct correctly.

Sub-disciplines now naming themselves

By 2026, two practitioner talks at AI Engineer London named the activity at finer granularity:

Both nest under agentic engineering. Karpathy's framing is the umbrella ("how do we go faster, properly?"); Lopopolo and Debois name which knobs you turn to do that.

Reaches CS education

Raymond Fu (Learning Software Engineering During the Era of AI (Raymond Fu, TEDxCSTU), May 2025) translates the same floor/ceiling split into a curriculum for undergrads: "AI is raising the floor, but software engineers are raising the ceiling." His proposed curriculum (foundations → systems thinking → full-stack → cross-discipline → AI collaboration → adaptability) is essentially "how to grow into an agentic engineer from scratch," in language a CS-major's parent would buy. Signal: the framing isn't industry-only anymore.

2026-06-13 — A concrete 7-skill competency model

The 7 Skills You Need to Build AI Agents (IBM Technology) (Bri Kopecki) supplies the missing competency model for the prompt-engineering → agent-engineering shift this page has been circling. Her thesis is the same as Karpathy's umbrella ("go faster, properly"), but operationalised into a hiring-ready taxonomy: "prompt engineer" is a category error, and building production agents is a systems-engineering discipline. The seven skills:

  1. System design — agent as orchestra (LLM + tools + state + sub-agents); data flow and failure coordination → Architecture Engineering.
  2. Tool & contract design — every tool is a typed contract; strict schemas over "fill the gaps with imagination."
  3. Retrieval engineering — RAG quality caps agent performance; chunking, embeddings, re-ranking → Context Engineering.
  4. Reliability engineering — retries, timeouts, fallbacks, circuit breakers — the classic backend playbook.
  5. Security & safety — the agent is an attack surface; defend against Prompt Injection; least-privilege + approval gates.
  6. Evaluation & observability — tracing + eval pipelines; "vibes don't scale, metrics do" → Binary Eval Assertions, LLM as Judge.
  7. Product thinking — UX for unpredictable systems; confidence signalling, graceful failure, escalation → Human in the Loop.

The load-bearing claim: most of these are re-pointed backend / distributed-systems / security competencies, not net-new ones. The gap between a demo agent and a production agent is engineering rigor, not better prompts — which is precisely the floor/ceiling distinction this page opens with. Kopecki's closing line ("the prompt engineer got us here; the agent engineer will take us forward") names the same role transition.

2026-06-21 — The organisational scaffolding underneath (Zehren)

How Agentic AI Is Changing the CIO Role (Forbes, Tim Keary) adds the executive-organisational complement to the engineering discipline this page describes. Tom Zehren (CEO, Info-Tech Research Group) frames the C-suite need as a "scalable agentic AI framework" — the org-level scaffolding (governance, mandates, posture, capability) required for the engineering rigor Karpathy / Lopopolo / Debois / Kopecki call for at the code layer to compound across an enterprise. Without it, you have a few brilliant agentic engineers and an organisation that still says no. Zehren's positive-framed name for the role this asks of the CIO: Exponential IT.

Sources