Start Small, But Start Now

Deciding where to begin with AI coding can be daunting. Human practices typically do not evolve as fast as technology, so I believe the most important step is to simply start experimenting. We see a lot of commentary predicting that AI will be the end of human developers, but in my view, engineers aren’t being replaced by AI—they’re being empowered. In my opinion, AI coding tools are, and should be seen as, productivity enhancers that help developers build more solutions faster, and with fewer defects. From my perspective, the key is starting with a mindset of “there has to be a better way” and staying open to new techniques.

Rather than diving into the full "vibe coding" trend—where entire applications are built through prompts alone—I advise teams to start by applying AI to targeted, repetitive tasks. I believe writing first-draft code, researching APIs, translating code between languages, generating documentation, spinning up prototypes and composing test cases are all strong entry points.

Regardless of how it’s generated, I believe it’s critical for developers to read and deeply internalize the flow of their code before committing or releasing it. This isn't just a best practice; it’s a safeguard. AI-generated code can often be the least secure code in a system, making secure review processes, scanning tools and proactive cybersecurity measures even more important in any AI coding workflow.

Promoting an AI-Forward Team Culture

Tools alone typically don’t change how teams work; culture can play a significant role. The companies I’ve seen get the most benefit from AI coding to date are those that intentionally socialize various tools with their software teams and create an environment that encourages experimentation and knowledge-sharing. I work with one company that made Cursor—a popular AI coding integrated development environment—standard across its engineering organization. But it didn’t stop at the mandate. The team also maintains an active Cursor Slack channel to share successful practices and recognize developers who make the most of new capabilities. In addition, they hold regular Cursor meetups for enthusiastic engineers to exchange techniques, which are recorded and disseminated to the broader team.

Drive Better Outputs with Better Input

A common source of frustration with AI-assisted development is the perception that the tools don’t "get it." But frequently, the issue isn’t with the model—it’s with the input. in my view, developers should treat an LLM like a junior engineer who can write code but lacks architectural context, product intuition and domain knowledge.

To get useful results, I encourage developers to supply clear exposition, including goals, constraints, libraries in use, performance requirements and feature definitions. Visual aids like screenshots and wireframes can be surprisingly effective, especially during debugging. The good news: LLMs typically don’t need polished writing. Bullet points, snippets or partial documentation often work just as well. Many AI coding tools now support configurations that preload relevant documentation and standards into every prompt session, helping to ensure that outputs reflect how your team actually builds software.

AI Coding at Work: Small Shifts, Real Impact

Don’t be discouraged by a lack of immediate, measurable productivity improvements when rolling out AI coding tools. Software development is a creative process, and adapting to new workflows—especially those powered by AI—can take time. What may feel incremental at first often builds toward more durable, long-term gains.

Here are three examples of where I see some of those gains starting to show:

• Faster prototyping: What once took weeks to validate can now be built—and discarded—in a matter of hours. This lowers the cost of experimentation, encourages iteration and helps teams apply lessons learned earlier in the development cycle.
  ‍
• Interactive design exploration: During the design phase, developers can use LLM-powered AI chatbots to investigate implementation options far more efficiently than with traditional search. Instead of skimming documentation or scanning forums, they can ask targeted questions—such as which HTTP libraries are most commonly used in Python—and prompt the model to generate code snippets that demonstrate key differences. What once could take hours can now take minutes.
  ‍
• Streamlined test development: LLMs can eliminate one of the most tedious parts of the development process—writing repetitive test cases. When seeded with clear input, output and intent, they can generate reliable tests that support CI/CD workflows, reinforce code quality and help developers stay focused on higher-leverage work.

These improvements may not always show up in dashboards or metrics. But collectively, they represent a shift in how engineering teams can work—enabling faster validation, better decision-making and more consistent delivery.

Scale Impact with Agentic Workflows

“Agentic AI” refers to the next generation of AI systems: tools that can perform multistep tasks with minimal human input. These agents are given objectives, context and access to the systems they need to operate—and they're evolving quickly. While promising, I believe they work best when introduced after teams have already gained comfort with more basic AI coding assistants.

To use agentic AI effectively, I recommend that teams begin by defining small, discrete tasks. Breaking down objectives into modular components not only helps AI agents work more effectively but can also make their output easier for engineers to validate and debug. This modular approach also makes it possible to run multiple agents in parallel on tasks like testing, documentation, code generation and analysis.

That added power comes with increased risk. Generative AI is advancing quickly but is still prone to errors. In my view, frequent version control check-ins are essential, making it easy to revert to a known-good state if something breaks. As with any AI-generated code, a knowledgeable human should review and fully understand the changes before committing to a codebase.

I believe teams that are already practicing modular development, writing clear specifications and investing in strong documentation will be best equipped to take advantage of these more autonomous tools.