TL;DR
Google’s May 2026 whitepaper, The New SDLC With Vibe Coding, says software teams are moving from hand-writing code toward directing AI agents. The paper reports broad use of coding agents and argues the model is only a small part of agent behavior, while harnesses, tests, evals and engineering discipline carry most of the work.
Google’s May 2026 whitepaper The New SDLC With Vibe Coding argues that software engineering’s main AI shift is not model selection but the discipline around AI agents, a claim that matters as the paper says 85% of professional developers regularly use coding agents and about 41% of new code is AI-generated.
The paper, by Addy Osmani, Shubham Saboo and Sokratis Kartakis, frames AI-assisted development as a spectrum. At one end is casual vibe coding: brief prompts, light review and code that may be useful for prototypes or disposable scripts. At the other is agentic engineering, where AI output is governed by formal specs, automated tests, evals, CI gates and human architectural review.
The paper reports that 51% of professional developers use AI coding agents daily. It also argues that the line between low-risk and high-risk AI development is verification, not the presence of AI itself. Tests check deterministic behavior, while evals check agent choices, tool use and quality thresholds for less predictable work.
The strongest claim is the model-harness split. The paper describes an agent as the model plus the surrounding harness: prompts, tools, context, hooks, sandboxes, sub-agents and observability. It puts the model at roughly 10% of behavior and the harness at roughly 90%, citing cases where benchmark results improved after teams changed the harness while keeping the model fixed.
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 for engineering leaders because it shifts budget and management attention away from model shopping alone. If agent behavior is shaped mainly by scaffolding, then a team’s real advantage may sit in its specs, test coverage, eval design, context policies and routing choices.
It also reframes AI coding costs. Thorsten Meyer AI’s analysis reads the paper as a warning that casual AI use can look cheap while creating higher operating costs through token-heavy repair loops, maintenance work and security fixes. The paper also cites METR findings that some AI-assisted tasks took 19% longer, underscoring that verification can erase expected gains when workflows are weak.
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.
Vibe Coding Gets Narrower
The whitepaper builds on Andrej Karpathy’s February 2025 phrase ‘vibe coding,’ which described accepting AI-generated code with minimal review and pasting errors back until something worked. The term later spread across many AI-assisted workflows, making it less precise.
Google’s paper narrows the term by placing vibe coding at the risky end of a broader spectrum. In that framing, production work should move toward agentic engineering, a more structured process that treats AI as an implementation engine rather than an unchecked author of software.
“generation is solved; verification, judgment, and direction are the new craft”
— Osmani, Saboo and Kartakis, Google whitepaper
Benchmarks Need Real-World Tests
Several points remain unsettled. The 10% model and 90% harness split is presented as a rough strategic framing, not a universal measurement across all teams or tasks. The source material also does not provide enough detail to judge the survey design behind the reported 85% adoption figure.
It is also not yet clear how widely the cited benchmark gains will carry over to large, messy production codebases. Thorsten Meyer AI describes the paper as mostly tool-agnostic but also says the on-ramps point toward Google’s own Gemini, Jules and Agent Development Kit products, so readers should separate the workflow claims from vendor positioning.
Teams Test The Harness
The next phase is practical adoption. Teams weighing AI coding agents will likely compare not only models, but also test depth, eval design, CI gates, tool permissions, context quality, observability and model-routing policies.
For Google, the paper gives a framework for selling agentic development around its own products. For readers, the immediate question is whether their current AI workflow can prove correctness and manage cost before more generated code reaches production.
Key Questions
Is Google saying models do not matter?
No. The paper argues that models matter, but that agent behavior is shaped heavily by the surrounding harness, including prompts, tools, context, tests, evals and safeguards.
What does vibe coding mean here?
In this framing, vibe coding means casual AI coding with light review and weak verification. The paper treats it as one end of a spectrum, not as a label for all AI-assisted development.
What is a harness in AI coding?
A harness is the operating layer around the model: instructions, tool access, context rules, sandboxes, hooks, sub-agents, logging and feedback systems that shape what the agent does.
What remains unproven?
The exact model-versus-harness split is not settled, and the cited benchmark gains may not map cleanly to every production environment. Adoption data also needs source-level review before being treated as an industry-wide fact.
What should engineering teams watch next?
Teams should watch whether agentic workflows reduce rework, security fixes and token spend in real projects. The key test is whether AI-generated code can pass clear verification standards before it ships.
Source: Thorsten Meyer AI
