When Code Is Free, Proof Becomes Your Only Leverage

At AI Engineer Europe, Ryan LeFebvre urged skipping code, while Mario Zechner insisted on reading every critical line. Alex Volkov reframes it not as personality but as a task continuum—the same engineer must move fluidly between extremes. The real question is proof per change.

Evidence supports the tension: GitHub is on track for 14 billion commits (14x YoY), Faros AI’s survey shows an 861% increase in code deletion per PR and a 242% jump in incidents per PR. Anthropic ships 8x more code but warns that human review has become the bottleneck. Speed is pulling quality downward.

Volkov distills advice from leading AI engineers into a routing table. Read every line for auth, payments, permissions, and irreversible data. Decompose large PRs, separate writing from reviewing agents, and use traces, evals, and shadow mode. Engineer linters and guardrails so human attention catches recurring mistake patterns once, not every time.

Capability drift is moving the proof layer upward. Newer models like Mythos tempt engineers to stop looking at code entirely, but self-verifying loops hide review without removing it. Leadership’s job is to design proof systems that adapt as model ability grows, ensuring judgment stays where risk demands it.