TL;DR
A new Google whitepaper, “The New SDLC With Vibe Coding,” argues that software development is moving from direct code writing toward intent-driven work with AI agents. The paper says the model itself is only a small part of agent performance, while tests, evals, tooling, context and oversight now determine whether AI-generated code is usable in production.
Google has published a new whitepaper that argues software engineering is being reshaped less by any single AI model than by the systems used to direct, test and verify AI-generated code, a claim with direct consequences for engineering teams already using coding agents at scale.
The paper, The New SDLC With Vibe Coding, was written by Addy Osmani, Shubham Saboo and Sokratis Kartakis. According to the paper, 85% of professional developers regularly use AI coding agents, 51% use them daily and roughly 41% of new code is AI-generated as of early 2026. Those figures are presented by the authors as evidence that AI-assisted coding is no longer a fringe workflow.
The paper’s central argument is that teams should focus less on model choice alone and more on the “harness” around the model: prompts, tools, context, rule files, hooks, sandboxes, tests, evals and observability. The authors frame a working agent as “model plus harness,” with the model accounting for roughly 10% of behavior and the surrounding system accounting for the rest.
The paper separates casual “vibe coding” from more disciplined “agentic engineering.” In the authors’ framing, vibe coding relies on prompts, quick checks and repeated fix loops, while agentic engineering uses formal specifications, automated tests, evals, CI gates and human architectural review. The dividing line, the paper argues, is verification.
The model is only 10%
A Google whitepaper argues software’s biggest shift is from writing code to expressing intent. Its sharpest claim: the model you obsess over is the smallest part of the system — the scaffolding around it does the real work.
The clearest map yet of how serious AI development works — and mostly tool-agnostic. But it’s a Google funnel: the concepts are neutral, the on-ramps point to Gemini, Jules & the ADK. If the harness is 90% and it’s yours, your moat and your costs both live there — so own your scaffolding, route across models, and remember: AI amplifies whatever engineering culture it lands in.
Verification Becomes The Bottleneck
The paper matters because it shifts the business and engineering question from “which model is best?” to “what system makes AI output reliable?” If the authors’ framing holds, companies that treat AI coding as a prompt-only shortcut may face higher costs later through rework, fragile code and security remediation.
The paper also challenges how teams budget for AI development. It describes casual AI coding as low upfront cost but high operating cost, while disciplined agentic engineering requires more investment in specs, evals and context management before code is generated. The source material says the paper cites task slowdowns reported by METR, including a 19% increase on some tasks, as evidence that verification can be a real cost rather than a footnote.
For readers managing software teams, the practical point is that AI adoption alone does not guarantee faster or safer delivery. The paper’s claim is narrower: AI can amplify an engineering organization’s existing habits, including weak testing, unclear requirements or poor review discipline.
Coding with AI For Dummies (For Dummies: Learning Made Easy)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Vibe Coding To Agents
The term “vibe coding” was popularized by Andrej Karpathy in February 2025 to describe a loose style of AI-assisted programming in which a developer leans on generated code and feeds errors back to the model until something works. The Google paper treats that as one end of a spectrum rather than a label for all AI coding.
At the other end, the paper uses “agentic engineering” for workflows where AI agents operate within formal constraints. That includes written specs, automated tests for deterministic behavior, evals for less predictable agent behavior, CI/CD gates and human review for architecture and product judgment.
The source material also flags a commercial angle. While the concepts in the paper are described as tool-agnostic, the surrounding product path points toward Google’s AI developer stack, including Gemini, Jules and the Agent Development Kit. That means the paper can be read both as a technical framework and as a product strategy document.
“The biggest shift is from writing code to expressing intent.”
— Osmani, Saboo and Kartakis, Google whitepaper
Portable Mini Inductor Tester, Type-C Powered High Precision Mainboard Coil Testing Tool, Fast Inductance Fault Detection Diagnosis Repair Tool for Mobile Phone Electronic Components-2 Pcs
Instant Inductor– In-Circuit Friendly：Simply bring the sensing tip close to the target inductor directly on the board –…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Claims Still Need More Proof
Several claims remain dependent on the paper’s own framing and cited examples. The 10% model and 90% harness split is presented as a rough guide rather than a universal measurement. It is also not yet clear how consistently that ratio applies across small applications, large legacy systems, regulated industries or teams with limited testing infrastructure.
The benchmark examples are specific but not final proof of a broad rule. The source material says one team moved an agent from outside the Top 30 to the Top 5 on Terminal Bench 2.0 by changing only the harness, and that a LangChain experiment raised an agent score by 13.7 points through prompt, tool and middleware changes. Those examples support the paper’s argument, but they do not settle how much harness design matters in every production environment.
It also remains unclear how much of the reported AI-generated code share is accepted without major edits, how much survives long-term maintenance and how much creates later security or reliability work.
AI ESP32-P4-WIFI6-DEV-KIT Development Board, Based On ESP32-P4 and ESP32-C6, Supports Wi-Fi 6 and Bluetooth 5 / BLE 5, Rich Human-Machine Interfaces, Supports AI Speech Interaction, with Speaker
ESP32-P4-WIFI6-DEV-KIT Development Board, Based On ESP32-P4 and ESP32-C6. It features rich Human-Machine interfaces, including MIPI-CSI (with integrated Image…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Teams Face A Build Choice
The next step for engineering leaders is likely to be deciding how much of the AI coding stack they want to own. The paper’s argument points toward investment in shared prompts, internal tools, repository context, evals, test suites, sandboxes and model routing rather than relying only on a single provider’s model improvements.
Readers should watch for whether Google, other platform companies and open-source agent frameworks turn this harness layer into a new competitive market. More public results from production teams will also matter: the paper gives a framework, but broader evidence will determine whether its model-versus-harness split becomes standard engineering practice.
Supply Chain Software Security: AI, IoT, and Application Security
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the actual news development?
Google has published a May 2026 whitepaper, The New SDLC With Vibe Coding, arguing that AI-assisted software development is moving from code writing toward intent setting, agent direction and verification.
What does the paper mean by “the model is only 10%”?
The paper argues that a coding agent’s behavior depends mostly on the harness around the model: prompts, tools, context, tests, evals, sandboxes and observability. The 10% figure is a rough framing, not a settled industry measurement.
How does the paper distinguish vibe coding from agentic engineering?
Vibe coding is described as casual prompting with limited review, suited mainly to prototypes or disposable scripts. Agentic engineering adds formal specs, automated tests, evals, CI gates and human oversight for production work.
What is confirmed right now?
The confirmed development is the publication of the Google whitepaper and the claims it makes. The adoption figures, benchmark examples and cost arguments are attributed to the paper and the source analysis, not independently established here.
Why should software teams care?
If the paper’s argument is right, AI coding performance will depend heavily on engineering discipline around verification and context. Teams that skip that work may see short-term speed gains offset by rework, maintenance problems and security fixes.
Source: Thorsten Meyer AI
