Founder argues human code review finds only 30 percent of defects

Jonathan Ellis, founder of Brock AI, argues that human code review is no longer a cost-effective use of engineering time as AI code generation speeds up to 10x. Ellis cites research indicating that human code review typically identifies only 30% of defects, a figure that drops to 20% in complex domains like security. He contends that while humans have vivid memories of finding bugs, they are statistically poor at estimating the value of review time compared to other tasks. Ellis advocates shifting human effort from reading diffs to higher-level strategy, such as defining requirements via RFCs and architecture decision records, and building validation loops. He notes that agents can now handle long-running tasks, with Ellis running 12 parallel tasks on Fable for hours or days. He also highlights that agents are 'confidently wrong' in edge cases, citing an instance where Mjolnir, a Rust-based tool, cloned entire conversations due to poor memory allocation patterns. Despite these issues, Ellis believes the trend is improving. He argues that the industry must abandon the practice of humans reading code diffs for review, as the time invested yields diminishing returns compared to strategic planning and automated validation.