Essay 01/state as of July 2026

The Spectrum

"Vibe coding" was never one practice. By mid-2026 the field had sorted itself into four schools, separated by how much structure stands between an intention and a merge.

The label "vibe coding" entered the vocabulary as a shrug — accept whatever the model writes, run it, see. Andrej Karpathy coined it on February 2, 2025 Karpathy, Feb 2025; Collins Dictionary made it Word of the Year within ten months Collins, Nov 2025. By then the term was already splitting in two. Simon Willison proposed "vibe engineering" for professionals who accelerate with LLMs while staying accountable Willison, Oct 2025, and by February 2026 Karpathy himself had relabeled the disciplined end "agentic engineering" Karpathy, Feb 2026. The name is still contested; the distinction is not. Throwaway work and production work now run on visibly different practices.

§1The four schools

Sorting the practice by structure, not by tool, yields four schools. Each is legitimate somewhere; the failures come from using one school's habits at another school's stakes.

low structure high structure Vibe coding prompt → accept → run prototypes, spikes Vibe engineering review gates, tests around generation Spec‑driven spec → plan → tasks plan review is the code review Agentic orchestration fleets, harnesses, machine‑held gates
FIG 1 — the four schools, by how much structure stands between intention and merge. figure scrolls →

The taxonomy carries a decision rule. Complexity ceiling rises left to right — file, feature, system, ecosystem — and the verification method changes with it: a manual look, a prompt anchored in context, automated tests, agentic self-audit. The human role shifts in step, from intuitive curator to context architect to logic validator to goal orchestrator. What never changes is that some verification exists; the schools differ only in who or what performs it.

§2Harness engineering

The strongest articulation of the high-structure end arrived as a case study. A small OpenAI team reported building and shipping a beta product of roughly a million lines over five months with no manually written source code OpenAI, Feb 2026. Its one-line thesis: "the discipline shows up more in the scaffolding rather than the code." The practices are concrete — the app boots per git worktree, agents query their own logs and metrics, linters carry remediation text written for agents, a doc-gardening agent tends the knowledge base, and an agent that struggles is read as a signal of a missing tool or guardrail, fixed by having the agent write the fix.

The report also names the cost. As generation throughput rose, the team's binding constraint became human QA capacity — verification, not generation OpenAI, Feb 2026. Martin Fowler wrote the practice up Fowler, Apr 2026, and "harness engineering" entered the Thoughtworks Technology Radar, framed as the infrastructure, constraints, and feedback loops wrapped around agents Radar Vol 34, Apr 2026. The same Radar makes a quieter, sharper point: "spec-driven development" and "harness engineering" are used inconsistently and overlap — possibly two names for one idea, encoded once as documents and once as executable environment.

One supply-side change made the harness the story of 2026: frontier models largely converged. When every agent runs a comparable model, the scaffolding around it does the differentiating work State of Vibe Coding 2026, synthesis — which is the scaffolding thesis restated from the other direction.

§3What the numbers do and don't say

A cluster of statistics circulates with this story, and most of it deserves flags rather than confidence. That roughly 84% of professional developers use AI tools 2026 roundups contested · methodology uneven is directionally safe. Two more specific claims travel widely on thinner legs: a mid-2025 randomized trial found experienced developers about 19% slower with AI on familiar codebases METR, 2025 contested · single study, and an application-security vendor reported around 45% of AI-generated code carrying an OWASP Top 10 flaw Veracode, 2025 vendor claim. Real studies exist behind each, but the repetition outruns the methodology. Treat them as cautions, not measurements.

§4Where vibe coding stays legitimate

Nothing in the mature practice bans the loose end of the spectrum. Prototypes, spikes, and UI polish are exactly where accept-and-run iteration earns its keep — ill-defined, high-creative, low-risk work where the cost of being wrong is a discarded branch. The recurring failure mode is vibe-coding habits carried into production stakes, or, at the 2026 pendulum's other extreme, heavyweight spec ceremony applied to a throwaway. Healthy teams oscillate deliberately. The counter-demonstration is worth keeping in view: François Zaninotto built a working 3D sculpting tool in about ten hours of incremental natural-language iteration, with no formal spec at all Zaninotto, Nov 2025.

§5The casualty ledger

The spectrum has a labor-market shadow. Software-development employment for ages 22–25 has fallen roughly 20% since 2022 Stanford AI Index 2026 — the routine work that trained a generation is the work the machines absorbed first. Steeper figures circulate for entry-level postings, but they vary by source and definition and remain unverified unverified aggregations. The structural question — who learns judgment when the apprenticeship is automated — has no answer in the mid-2026 record. Essay 06 carries it forward as an open problem.

§6Sources