Anthropic Challenges Traditional Mental Models in AI-Driven Code Review
Anthropic shifts code review to AI-first workflows, questioning the necessity of human mental models.
1Scaling Through Automation
Anthropic transitioned to an AI-first code review process after their monorepo volume surged to 30,000 daily changes by September 2026. The team now relies on Claude to write 90% of Claude Code itself. This shift aims to ensure most pull requests merge with at most one human participant, maintaining stability without increasing incident rates.
2Abandoning Mental Models
The engineering team claims that no individual needs a complete mental model of the codebase. Instead, developers request Claude to generate explainers for specific code slices on demand. Critics argue this removes the ability to perform manual, causal debugging, creating a dependency where the system relies entirely on the model's ability to interpret itself.
3Operational Friction Costs
Automated agents create unique overheads, such as excessive review rounds. One CLI pull request underwent over fifty iterations before intervention. The team discovered that repeated cycles often flagged existing code rather than new bugs. Consequently, they implemented caps on review rounds to prevent agents from endlessly editing lines based on incorrect internal state representations.
4Risk-Based Merge Gates
Merge approvals are now governed by a classifier-based risk tier system. Low-risk changes, such as documentation or additive updates, proceed without human oversight. Medium-risk changes require one human reviewer. Crucially, the system mandates that an agent's approval never replaces a human, and the original author counts as the required human reviewer for their changes.
5Issue Lifecycle Discrepancies
The repository's automated issue management has faced scrutiny regarding accuracy. Developer marcindulak identified over 6,000 closed issues that were marked inactive despite containing valid reproduction steps. Although Claude generated an explanation for the closure policy, the underlying workflow remains rigid, locking issues after seven days and preventing manual reopening by users or contributors.
6Compliance and Security Gaps
The current system lacks circuit breakers to tighten approval thresholds when metrics deviate. Furthermore, the security review GitHub Action is not hardened against prompt injection, limiting its use to trusted pull requests. Compliance requirements, such as SOC 2 mandates for dual-human review, remain unaddressed in the team's current technical documentation and public strategy.
7Industry Perspective on Complexity
Debates on Hacker News highlight that large legacy systems often lack comprehensive human mental models, suggesting this approach might be an upgrade. Proponents argue that replacing human review, which consumes up to one-third of engineering time, with AI-driven analysis provides significant cost savings and operational efficiency for teams managing massive, high-velocity codebases.
8Future of Autonomous Engineering
The shift at Anthropic signals a move toward agentic software development where human oversight is selectively applied. While the team successfully manages high change volumes, the reliance on AI for both implementation and review introduces new failure modes. Future success depends on implementing robust circuit breakers and addressing critical gaps in auditability.
Original article: InfoQ — Anthropic Challenges Traditional Mental Models in AI-Driven Code Review