Anthropic reports that Claude now writes roughly 90% of the changes in its codebase, and the team no longer maintains a full mental model of the entire repository. The shift has produced pull requests that required more than fifty review rounds in a single night, with reviewers spending most of their effort on legacy sections rather than the AI‑generated changes.
AI‑generated code review at scale
The volume of review iterations indicates that existing manual review processes may not scale when the majority of code originates from an LLM. Practitioners should anticipate longer review cycles and consider augmenting human review with automated linting, static analysis, and test generation to keep throughput acceptable.
Implications of a missing mental model
Without a comprehensive mental picture of the codebase, engineers may struggle to assess the impact of new changes, increasing the risk of regressions or security oversights. This situation suggests a need for stronger documentation, architecture diagrams, and tooling that surface code ownership and dependency maps.
Operational and security considerations
Relying heavily on AI‑generated code raises several practical concerns:
- Verification of intent: reviewers must confirm that generated snippets align with design intent and do not introduce hidden logic.
- Testing coverage: extensive unit, integration, and fuzz testing become critical to catch defects that may not be obvious in review.
- Auditability: maintaining logs of AI prompts and generated outputs can help trace the origin of a change during incident investigations.
- Governance gaps: the team acknowledges a missing control for tracking the mental model, highlighting a governance area that needs formalization.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Engineers should evaluate their current review pipelines for automation opportunities, reinforce documentation practices, and establish explicit controls for tracking AI‑generated contributions. Monitoring test results, audit logs, and dependency maps will help mitigate the risks introduced by a codebase that is largely authored by an LLM.