As a specialist in enterprise SaaS technology, Vijay Raina has spent his career at the intersection of software architecture and team leadership. He’s now focused on the next great shift in development: the integration of artificial intelligence. In our conversation, we explore the starkly uneven productivity gains from AI, the emergence of a new “AI coach” role, and the deep-seated challenges engineering teams face as they navigate a future where their relationship with code is fundamentally changing. We delve into strategies for effective AI-assisted planning, the messy reality of “vibe coding,” and the critical need to restore a sense of craftsmanship and joy for developers feeling more like factory managers than artisans.
Some studies suggest experienced developers can be up to 19% slower when using AI. Why is this productivity gain so uneven across the industry, and what specific challenges do developers face when applying AI to legacy enterprise codebases compared to greenfield projects? Please share an example.
That 19% figure is really misleading because it averages out a massive spectrum of experiences. The reality of AI’s impact is incredibly uneven. On one end, you have small, cutting-edge teams working on greenfield projects with modern technologies like Node, Python, and React. For them, AI is a rocket ship; they can hit maximum speed, seeing productivity gains of a thousand percent. But that’s just not the world most developers live in. The other end of the spectrum is the vast landscape of enterprise development, filled with legacy code. The large language models simply weren’t trained on your company’s internal libraries or the ancient versions of frameworks you’re still forced to maintain. You can’t just rewrite the world’s code overnight, so for those developers, AI tools often feel completely useless, alienating them from the hype they hear from the other side.
You’ve noted the emergence of a new developer role that focuses more on tweaking prompts and building subagents than writing code. Could you detail the key skills this “AI coach” needs, and how this differs from the responsibilities of a traditional senior developer or team lead?
Absolutely, we’re seeing this fascinating new role emerge that doesn’t have a clear name yet, but I call it the “AI coach.” It’s not quite a developer and not quite a manager; it’s something in between. These individuals spend their time tweaking the factory, not the code itself. When an AI generates bad code, a traditional senior dev would dive in and fix it. The AI coach, however, fixes the prompt, adjusts a rules file, or even builds a dedicated subagent to prevent that error from happening again. Their core responsibility is to refine the AI system, to teach it and guide it. A traditional team lead manages people and process, whereas this role manages the AI’s behavior and performance, essentially hacking together the pieces of an intelligent development system that doesn’t yet exist off the shelf.
Given the importance of upfront planning, what are the essential steps in an effective AI-assisted planning process? Explain how using distinct AI personas, such as a requirements gatherer and a software architect, can help prevent common errors and reduce long-term technical debt.
Effective AI-assisted planning requires moving beyond a single tool and thinking in terms of a collaborative team of agents. We see this process as having at least three key steps, each handled by a distinct AI persona. First, you need a requirements gatherer that acts like a product manager, asking critical questions: Who is this feature for? What does it need to do? Then, you need a software architect persona to handle the technical requirements—deciding on frameworks, libraries, and architecture. Separating these concerns is crucial. This approach mimics the collaboration of a high-functioning human team, where a product manager, designer, and engineer bring different expertise. By having these AI personas sanity-check your decisions early on, you catch gaps and flawed assumptions before a single line of code is written, which is fundamental to preventing that long-term, painful technical debt.
The concept of “vibe coding” allows for building at the speed of thought but often results in messy, unmaintainable code. What specific, automated maintenance tasks should AI agents handle to clean this up, and what metrics can we use to ensure they refactor for quality?
“Vibe coding” is this magical experience where you can build at the speed of thought, but the crash afterward is brutal. You’re left with a 2,000-line file that’s a complete nightmare. My perfect world is one where humans handle the creative, in-the-moment vibe coding, and AI handles the cleanup. The AI agents should be responsible for the boring, mundane work that no one enjoys: refactoring massive files into proper modules, upgrading dependencies like a painful Python 2 to Python 3 migration, keeping libraries up-to-date, and filling in the gaps in unit test coverage. To ensure quality, we can use metrics like code complexity, test coverage percentages, and adherence to style guides. The ultimate goal is to allow developers to merge their creative “rubbish” and have an AI agent sweep up behind them, ensuring the codebase remains clean, maintainable, and robust without killing the creative flow.
Many developers feel a loss of craftsmanship, comparing their new role to managing a factory rather than being an artisan. How can engineering leaders help their teams navigate this “grief,” and what strategies can they use to restore a sense of joy and purpose in this new workflow?
This is a very real and painful transition for many. I’ve spoken with developers who are genuinely going through the five stages of grief—from denial that AI is here to stay, to anger that they’re being forced to use tools that produce “crap.” One developer told me he felt he’d gone from being a craftsman whittling a perfect chair to being a factory manager at Ikea, just shipping low-quality furniture. To navigate this, leaders must acknowledge this sense of loss. They need to reframe the narrative. The goal isn’t to replace craftsmanship but to elevate it. We can restore joy by automating the drudgery—the dependency upgrades, the boilerplate code, the endless refactoring—and freeing up developers to focus on creative, complex problem-solving. This allows them to stay in that state of flow where the real magic happens, bringing purpose back by letting them tackle the most interesting challenges, not just review thousands of lines of machine-generated code.
Current AI tools often focus on individual productivity, which can sometimes disrupt team collaboration. What does a high-performing human-AI software team look like in practice, and what specific features would an AI need to build “team flow” and act as a knowledgeable, integrated team member?
This is a critical point. So much of the focus is on the lone genius developer, but truly high-performance software development comes from great teams. I learned that if you want great products, you build great teams. A high-performing human-AI team is one where the AI is not just a tool but an active, knowledgeable member. It’s not about an individual prompting, waiting, and getting snapped out of their flow. It’s about creating that energy you feel in a room when everyone is on the same page, bouncing ideas off each other. To achieve this, an AI needs memory. It needs to learn from every interaction, understand who’s who on the team, what their expertise is, and tailor its suggestions accordingly. It needs to become an employee of the organization that helps bridge people together, answers questions, and shares social knowledge, not a day contractor who shows up with no context, makes a mess, and leaves.
What is your forecast for AI in software development? In the next three years, what will be the primary role of a human developer, and what core capabilities will the most advanced AI coding agents possess?
In the next three years, I believe the primary role of a human developer will shift significantly toward creative problem-solving, system design, and acting as the conductor of an orchestra of AI agents. The core skills will be problem decomposition and strategic thinking, not line-by-line coding. The most advanced AI agents will have moved far beyond the simple chat boxes we see today. They will be proactive, not reactive. Imagine an agent that doesn’t just wait for your command but identifies potential issues in your codebase, suggests refactoring opportunities, or even invents its own work based on new library releases or security vulnerabilities. These agents will possess deep organizational context, acting as a persistent, learning member of the team that can onboard new engineers, maintain code quality automatically, and accelerate the entire development lifecycle from planning to maintenance. The human’s job will be to provide the vision and the creative spark, while the AI handles the execution and upkeep with intelligence and autonomy.
