Agentic engineering is 📝Andrej Karpathy's February 2026 framing for the disciplined successor to 📝vibe coding — the practice of orchestrating AI coding agents as the default mode of software development, with the human acting as oversight rather than line-by-line author.
The term was introduced in Karpathy's one-year retrospective on vibe coding and carries two intentional load-bearing words. Agentic names the shift in default: writing code directly is no longer the baseline; orchestrating agents is. The human's primary surface becomes the prompt, the context window, the toolchain, and the review loop. Engineering insists this is a learnable discipline with real depth — not a vibe, not a workflow trick, not a temporary stage on the way back to "real" programming. Where vibe coding described a permissive style appropriate for prototypes, agentic engineering names the production-grade practice that emerges when the same default extends to systems people are expected to ship and maintain.
The framing implicitly resolves a critique vibe coding attracted — that it encouraged sloppy, unreviewed AI output — by reasserting standards inside an agent-first workflow rather than retreating from agents. It has been adopted quickly across the AI-assisted development discourse and now anchors discussions of how to operationalize coding agents at scale, including stacks built around Claude Code, Codex, and the agent-orchestration patterns Karpathy enumerated in his December 2025 "never felt so behind" tweet.
Agentic engineering is the framing I use when explaining why MythOS is infrastructure rather than a notes app. The shift from writing code to orchestrating agents only compounds if the agents have a durable knowledge substrate — which is exactly the gap MythOS fills.