Google recently released a comprehensive 51-page master class on AI coding, outlining how the industry is converging on best practices for the AI-driven software development life cycle (SDLC). This summary synthesizes the core frameworks presented, focusing on the shift from manual coding to system orchestration and why the surrounding infrastructure matters far more than the model itself.

The Shift to an AI-Driven SDLC

In a traditional SDLC, engineers spend the majority of their time implementing code. In an AI-driven SDLC, implementation is compressed from weeks to minutes or hours using AI agents. However, this shifts the development bottleneck to the beginning (requirements gathering) and the end (validation and testing) of the cycle. Business output is often limited not by how fast code is written, but by how quickly specifications are defined and validated.

The AI Coding Spectrum

AI coding is not binary; it exists on a spectrum of sophistication depending on the task’s requirements:

• Vibe Coding: Writing casual, high-level prompts with minimal planning and informal verification. While high-risk, it is useful for quick MVPs or disposable code.
• Structured AI-Assisted: Utilizing more detailed prompts and manual spot-checking of the output.
• Agentic Engineering: Creating highly engineered specifications, workflows, and automated evaluations to allow agents to autonomously iterate, debug, and test their own code.

The Harness: The 90% Rule

A critical takeaway from Google’s guide is that the LLM model itself only accounts for about 10% of a coding system’s effectiveness. The remaining 90% relies on the ‘harness’—the custom layer of rules, tools, context, guardrails, and orchestration built around the model. Investing in a robust harness allows standard models to perform at elite levels, making custom platform engineering far more valuable than simply waiting for better underlying LLMs.

Static vs. Dynamic Context

Managing an LLM’s limited context window is vital to avoid high costs and ‘context rot’. Google delineates context into two types:

• Static Context: Fixed rules and system prompts loaded into every session. It ensures reliability but is expensive to scale.
• Dynamic Context: Specialized skills and codebase conventions loaded on-demand. By leveraging dynamic context, teams can use a single generalist agent that dynamically scales into specialized roles rather than maintaining complex, fragile multi-agent systems.

The Economics of AI Coding

While ‘vibe coding’ has low initial Capital Expenditure (CapEx), it results in high Operational Expenditure (OpEx) due to massive token consumption spent correcting low-quality code. Conversely, Agentic Engineering requires high CapEx upfront to build the harness, but yields low OpEx and 3 to 10 times greater reliability over time.

Mentoring question

Looking at your current development workflow, are you mostly ‘vibe coding’ or are you investing in building a ‘harness’ to shift toward true agentic engineering? What is one small step you can take today to systematically capture your coding rules and workflows?

Source: https://youtube.com/watch?v=zbmuiaPuiNM&is=hK4atIJhPl4pnnG5