Serge Haziyev, CTO, Advanced Technologies at SoftServe. 25+ years helping Fortune 100s and startups capitalize on emerging tech.
The vibe coding phenomenon emerged earlier this year—seen by some as a joke, by others as a threat and by a few as a genuinely helpful productivity approach. Ever since Andrej Karpathy introduced the term to the community, it has sparked ongoing debate across social platforms.
But one key question remains: Can this prompt-driven coding style, enabled by agentic AI tools like Cursor and others, scale across development teams and be harnessed by businesses investing heavily in AI for productivity?
While vibe coding proves valuable for generating small applications and rapid prototypes, it struggles when applied to larger codebases and more complex software systems developed by teams rather than individuals. The challenge often lies in the fact that the resulting code falls far short of production-grade quality.
An increasing number of practitioners, myself included, believe that to effectively leverage agentic AI in real-life projects, we need to redefine the software development paradigm, particularly the people, processes and tools components of the SDLC (or PDLC) equation.
For short, we can call this “agentic engineering” and define its key objective as achieving faster time to market with smaller teams and reduced manual effort, effectively lowering development costs without compromising the quality of the resulting software.
Let’s explore the three aspects of the SDLC that require revision: people, process and tools.
People
Most software development teams today follow the two-pizza team model, based on Jeff Bezos’ idea that no team should be so large that it takes more than two pizzas to feed. This typically translates to eight to 12 individuals, with about half of them spending their full day working directly with code.
Coding, in fact, is a highly brain-intensive activity. It’s no secret that in the software development industry, engineers often spend only about one hour per day actually writing code. Multiple studies from respected sources, such as Microsoft Research, support this observation.
On average, about three hours a day are spent working with code—reviewing, debugging and writing unit tests. The rest of the workday is often filled with communication, documentation or meditating on complex problems. The larger the team, the more time is typically spent on internal communication.
As AI agents can autonomously execute part of a developer’s activities, this leads to smaller teams where fewer people write code directly and instead leverage the agents, which act as a team of virtual developers.
Here’s something new in the one-pizza team model: a role called the Intelligence Engineer. This role is solely responsible for configuring, operating and customizing the agents. Think of it like a flight engineer in the early days of aviation—a crucial crew member responsible for the proper functioning of the aircraft.
Process
In traditional Scrum, which is commonly adopted across the software industry, each sprint lasts no longer than one month and typically spans two weeks. The cadence is defined by the time lag between planning and the retrospective.
Now, imagine for a moment that development time has been reduced to zero. Once a user story enters the backlog, it immediately materializes as code.
Of course, AI agentic development tools aren’t quite there yet. But current performance already enables the translation of well-described tasks into code in minutes rather than hours or days.
Understanding this leads us to redefine the SDLC process, as some have already observed that structured communication and evaluation are becoming the new bottlenecks.
This shift elevates the role of written specifications: first, to align humans; and second, to translate intentions clearly to agents, minimizing ambiguity in the resulting program.
Wait, wasn’t that exactly what the waterfall process did, with its emphasis on upfront specification? The very approach that Scrum later challenged?
Waterfall emphasized upfront specifications to reduce ambiguity in execution. In the AI-driven era, we’re seeing a similar need, but not for static specs. Instead, we need executable intent: interpretable, testable and traceable descriptions that replace the role of traditional source code.
When, in the near future, code can be generated in minutes from a sprint backlog based on new specifications, the purpose of sprint cadence will change dramatically. It will no longer govern the “time to write code,” but rather the “time to think, align and evaluate.”
Tools
The plethora of prompt-driven coding tools appearing each month puzzles many engineering leaders. Evaluating the available tools for specific project needs alone could require a dedicated team working full time on just this activity.
Then comes the build versus buy question: Can universal, publicly available tools like Cursor, Windsurf, Lovable, Builder.io, Devin and others be effective in a specific project environment based on X, Y and Z technologies? Or is it better to build custom agents tailored to project needs using agentic frameworks such as Claude Code, OpenAI Codex or Gemini CLI?
Only time will tell which approach will prevail, but one trend we can already observe is that prompt engineering (the core of vibe coding) is giving way to context engineering, which promotes a more systematic approach to feeding LLMs with complete and relevant information.
While some companies are already experimenting with the context engineering approach alongside spec-driven development for their in-house agents, we can expect such tools to become available to a broader market soon, with kiro.dev being a good example.
Sure, bringing in context and project specifics requires much more effort than simply installing an IDE. But that’s the point with humans, too—think of it as onboarding a new team member who doesn’t drink coffee…or eat pizza.
The Future Of Software Development Starts With Redefinition
Vibe coding is uncovering the need for a systemic shift in how we build software. As agentic AI becomes more capable, organizations must move beyond the novelty of agentic engineering and embrace it as a disciplined framework. This means rethinking team composition, adapting processes to accelerate alignment and evaluation and choosing tools not just for output, but for orchestration.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
