AI Can Find the Code. It Didn't Know How the System Worked (www.wespiser.com)

A recent exploration highlights the limitations of AI in modifying large, complex codebases, particularly through the lens of Large Language Models (LLMs). The case involved attempting to add a simple UI feature within a sprawling monorepo, which included outdated code and lack of clear dependencies. While the AI could locate relevant files and generate code, it struggled to comprehend the overall system's behavior, leading to flawed implementations. This misalignment stems from the AI's failure to capture the non-obvious aspects of the software architecture, such as where integration points reside or how various components interact. This investigation is significant for the AI/ML community as it underscores a critical gap in the capabilities of current AI coding agents: understanding complex system architectures beyond mere code manipulation. Even advanced models like Claude Opus 4.7 showed only marginal improvements. The findings suggest that successful AI-assisted coding requires clearer documentation, architectural decision records, and onboarding processes to make the system legible for both AI and human programmers. Without addressing these foundational issues, the promise of rapid AI-driven development remains unfulfilled.