📊 Full opportunity report: The Model Is Only 10%: The Real Lesson of the New SDLC on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
A recent whitepaper highlights that in AI-assisted software development, the core challenge isn’t the AI model itself but the surrounding harness and context management. This shift impacts how teams should invest in AI tools and infrastructure.
A new Google whitepaper, The New SDLC With Vibe Coding, asserts that the most significant shift in software engineering isn’t a new language or framework, but moving from writing code to expressing intent and trusting machines to generate software. The paper states that the AI model itself accounts for only about 10% of the system’s behavior, with the remaining 90% determined by the harness—comprising prompts, tools, rules, and observability.
The whitepaper by Addy Osmani, Shubham Saboo, and Sokratis Kartakis highlights that a common misconception is to focus on the AI model’s capabilities, but in reality, the surrounding infrastructure—referred to as the harness—plays a far more significant role in system performance and reliability. Experiments cited show that tweaking the harness, such as prompts and tools, can drastically improve AI agent outcomes, often more than upgrading the model itself.
Furthermore, the paper introduces the concept of context engineering, emphasizing that the quality and management of information provided to the AI—through instructions, knowledge, memory, and guardrails—are critical. The authors argue that cost efficiency and system robustness depend heavily on how well this context is structured and maintained, rather than solely on the AI model’s raw power.
Finally, the whitepaper discusses the economic implications, noting that while vibe coding (minimal structure, quick prompts) appears cheap initially, it incurs high long-term costs through token burn, maintenance, and security vulnerabilities. Conversely, disciplined, agentic engineering—focused on building robust harnesses and context—offers a more sustainable approach, reducing operational costs over time.
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.
Implications for AI Development and Investment Strategies
This new understanding shifts the focus from chasing the latest AI model to investing in the infrastructure, tooling, and processes that shape AI behavior. Organizations that prioritize harness and context engineering can achieve better results, higher reliability, and lower costs, making this insight crucial for strategic planning in AI adoption.
The AI Prompt Playbook: Master AI Prompt Engineering with 140 Ready-to-Use Templates for ChatGPT, Claude, Gemini & Copilot
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Evolution of AI in Software Engineering
Historically, AI development emphasized improving models to enhance performance. However, recent experiments and industry reports, including this whitepaper, reveal that the surrounding infrastructure—prompts, tools, rules—has a greater impact on outcomes. This aligns with earlier trends where system design and configuration often outweighed raw algorithmic improvements in determining success.
As of early 2026, AI tools are now integral to software workflows, with over 85% of developers using AI agents regularly. The shift towards a focus on harness and context reflects an evolution from model-centric to system-centric AI engineering, emphasizing control, verification, and cost management.
“The behavior you experience in AI systems is dominated by scaffolding you can build, own, and improve—it’s not the model itself.”
— Addy Osmani
Unresolved Questions About Harness and Cost Savings
While the paper demonstrates that harness and context engineering are crucial, it remains unclear how organizations can best scale these practices across diverse teams and projects. The precise methods for measuring and optimizing harness complexity, and how to balance upfront investment with long-term savings, are still being developed.
Next Steps for AI-Driven SDLC Adoption
Organizations are expected to reevaluate their AI strategies, emphasizing the development of robust harnesses, context management, and verification processes. Future research may focus on standardized frameworks for context engineering and tools that simplify this process, aiming to reduce operational costs and improve reliability in AI systems.
Key Questions
Why is the AI model only 10% of the system’s behavior?
The whitepaper explains that the AI’s behavior is primarily shaped by the harness—prompts, tools, rules, and context—rather than the model itself, which only provides raw generation capabilities.
How can organizations improve their AI systems based on this insight?
Focusing on building and refining the harness and context management—such as better prompts, guardrails, and procedural knowledge—can lead to more reliable and cost-effective AI applications.
Does this mean models are less important?
Models remain vital for capabilities, but their influence on behavior is limited compared to the surrounding infrastructure. Prioritizing harness and context engineering offers a more strategic path to success.
What are the economic implications of this shift?
While vibe coding appears cheaper initially, it incurs higher long-term costs. Disciplined harness and context engineering, though more expensive upfront, reduces operational expenses and security risks over time.
What should I do now to adapt to this new understanding?
Invest in developing robust prompts, tools, and guardrails, and establish processes for managing context and verification—shifting focus from just model upgrades to system-wide engineering.
Source: ThorstenMeyerAI.com
