Why Human Teams Still Beat AI for Complex Code

AI coding assistants are great for many tasks, but engineers say they fall short on complex codebases where human collaboration excels.

In a recent InfoQ podcast, software engineers Asgaut Mjølne Söderbom and Ola Hast shared a counterintuitive finding from their experiments with modern development tools. After integrating AI assistants like Anthropic's Claude into their workflow, they concluded that while the technology is useful for many peripheral tasks, it is not yet ready for the core challenge of writing and refactoring code within complex, existing systems. These are often called “brownfield” projects—codebases with years of history, technical debt, and undocumented business logic. Instead of relying on AI, the engineers found more success and higher quality by doubling down on a deeply human-collaborative practice: mob programming. Their experience serves as a practical reality check on the current capabilities of AI in software engineering, suggesting that for the hardest problems, a team's shared understanding remains more powerful than an algorithm's generative speed.

Mob programming is a software development approach where an entire team works together on the same task, at the same time, on a single computer. One person, the “driver,” is at the keyboard, while the rest of the team, the “navigators,” guide the work, discuss the problem, and review the code as it is written. This method creates a high-bandwidth channel for communication and knowledge sharing. It is particularly effective for brownfield projects because the collective memory and diverse perspectives of the team are needed to safely navigate the code's complexities. An AI assistant, in contrast, lacks this deep, historical context. It can generate syntactically correct code but often fails to grasp the subtle interdependencies and unwritten rules of a mature system, risking the introduction of new bugs or the violation of established architectural patterns. The AI operates on a statistical model of code, whereas the mob operates on a shared mental model of the specific system.

This distinction is critical for developers, CTOs, and technical leaders deciding where to invest their resources. The allure of AI-driven productivity is strong, but misapplying these tools can be counterproductive. For greenfield projects or well-defined, isolated tasks like writing unit tests or generating boilerplate code, AI assistants can be a significant force multiplier. However, for the crucial and delicate work of modernizing a legacy system or adding a major feature to a core product, the risk of an AI making a subtle but critical error is high. The engineers’ findings suggest that a team’s time is better spent building shared context through practices like mob programming. This approach not only produces higher-quality code but also builds team cohesion, distributes knowledge, and reduces the risk of key-person dependencies, making the entire engineering organization more resilient.

The business implications extend beyond code quality to risk management and long-term strategy. Relying on AI for complex coding tasks in legacy systems can create a hidden form of technical debt. The code might work initially, but it may be brittle, hard to maintain, and poorly understood by the human team responsible for it long-term. In contrast, investing in collaborative human processes like mob programming is an investment in the team itself. It builds a sustainable and adaptable engineering culture capable of tackling the most difficult challenges. For founders and executives, the takeaway is that technology is not a substitute for a high-functioning team. The most significant productivity gains may not come from the latest AI tool, but from creating an environment where deep collaboration and shared ownership can thrive, especially when dealing with the valuable and complex software assets that power the business.

Looking ahead, the conversation is not about choosing between humans or AI, but about intelligently combining their strengths. Söderbom and Hast noted that AI was useful for “everything else” besides the core coding on their complex project, such as summarizing documents or drafting communications. The future of elite software development will likely be a hybrid model. AI will handle the routine, the repetitive, and the context-free, freeing up human engineers to focus their collective brainpower on architecture, system design, and collaborative problem-solving in high-stakes environments. As AI models become more sophisticated, their ability to understand context will improve, but for now, the most valuable asset in navigating a complex codebase is the shared consciousness of the human team that lives in it every day.

This provides a crucial, experience-based counterpoint to the hype around AI coding assistants. It shows that for complex, existing 'brownfield' projects, human collaboration methods like mob programming are still superior, helping leaders make smarter decisions about tooling and team structure.

Over-relying on AI for complex legacy code can introduce costly bugs and hidden technical debt. Investing in collaborative practices like mob programming, while seemingly slower, builds team knowledge and reduces long-term maintenance costs, leading to more resilient and higher-quality software.