What happened

Andrej Karpathy coined vibe coding in 2025; by early 2026 he spoke of agentic engineering: the question is not whether AI writes code, but how you govern what comes back. JetBrains data: 85% of developers use AI tools regularly; ~41% of new code is AI-generated.

Vibe coding treats verification as optional — “seems to work” fits prototypes, not production. Agentic engineering adds formal specs, rule files (AGENTS.md), CI, automated tests, and LM judges.

Context engineering ≠ prompt engineering: not “how to phrase the question” but “what a new engineer must know about the project” — stack, architecture, hard rules, workflow. Static context (rule files) is expensive in tokens; dynamic context (tool results, RAG) loads on demand.

Factory model: the developer’s primary output is not lines of code but the system that produces code. Two modes — conductor (hands-on in the IDE) and orchestrator (assign goals to agents, review results).

Why it matters

METR (2025) found experienced developers took 19% longer on complex tasks with early AI tools — while feeling 20% faster. The lesson is not “AI is useless” but “without structure, verification tax grows.” Teams with clear specs and review report +25–39% on implementation-heavy work; without review, bug rates rise up to 41%.

For leaders: rule files and eval suites are code — version control, PR review, named owners. A demo proves one success; an eval suite proves repeatability.

In practice

1. Rule file per project — ten lines: stack, architecture, hard rules, tests; extend after every agent mistake.
2. Tests before generation — tests specify intent better than prompts and give agents a feedback loop.
3. Review everything that ships — especially clever imports and edge cases.
4. Split prototype vs production — vibe for exploration, agentic for production.
5. Invest in scaffold — swapping models without AGENTS.md predictably breaks output.

Takeaway

Vibe coding is not a methodology — it is optional verification. Agentic engineering moves value into the scaffold: context, tests, guardrails. The 2026 developer often designs the assembly line, not every widget by hand — but specification discipline gets harder, not easier.