Thesis: AI-assisted code review is shifting from universal line-by-line human inspection toward layered verification that allocates human attention by risk. The company reports that merged pull requests to Claude Code repositories more than doubled from June to September, while the share later reverted fell by more than 40%.
Mechanism: Verification is distributed across pre-PR checks, automated PR reviews, security review and risk-tiered human approval, with feature flags, gradual rollout and rollback after merge. Why now: The team says agent-produced change volume has made manual line-by-line review difficult at its current pace, so it is reallocating human attention while guarding against added failure-response work. Its stated objective is to reduce human review time without increasing reverts or incidents.
First implication: Teams using code agents should evaluate review savings together with reverts, incidents and time spent fixing failures. Anthropic tracks approvals, comments and waiting time, with reverts and incidents as guardrails. Evidence boundary: These are company-reported results and process details for Claude Code; the post does not prove causal attribution or generalizability to other teams. First watch condition: Whether reduced human approvals persist without an increase in reverts or incidents as throughput grows. This is a distinct analytical record because no prior New Runtime record is supplied; it examines the verification mechanism and its guardrails rather than treating AI code review as a tool launch.