We Stopped CI, Abandoned Code Review, and Embraced AI Pair Programming (www.arcblock.io)

In a recent exploration of AI development methodologies, a team led by an engineer made significant shifts in their approach, notably abandoning continuous integration (CI) and manual code reviewing in favor of AI pair programming. They found that while large language models (LLMs) excel in code generation, their effectiveness diminishes without close human oversight. The team discovered that AI could drift from the original goals during autonomous tasks, leading to a state of misleadingly "green" tests that masked underlying issues. This realization prompted a shift to a "small-step iteration" approach where AI output is regularly reviewed, preventing errors from compounding and allowing for timely course corrections. This shift has profound implications for the AI/ML community, challenging traditional engineering practices that prioritize extensive pre-implementation planning. The engineer emphasized the need for a balance between AI's agile execution capabilities and human oversight to maintain project alignment with broader objectives. They found that adopting a test-driven development (TDD) approach was crucial, as AI struggles to generate meaningful tests when code is created first, often producing weak validations. By framing prototypes as vehicles for learning—rather than final products—the team demonstrated that iterative prototyping could dramatically enhance efficiency in AI-assisted projects, paving the way for future Agile practices in AI development.