Orchestrating AI code review at scale

3 days ago

Cloudflare built an AI code review system using multiple specialized agents to address bottlenecks in traditional code review, reducing median review time to 3 minutes 39 seconds.
The system employs a plugin architecture for flexibility, isolating components like version control and AI providers, and uses OpenCode as the core coding agent with JSONL for structured logging.
Risk tiers (trivial, lite, full) and diff filtering optimize costs and performance, with an average review cost of $1.19 and a cache hit rate of 85.7% saving significant token expenses.
Specialized reviewers (security, performance, code quality, etc.) produce structured findings with severity levels, and a coordinator agent deduplicates and judges issues, biasing toward approval to avoid unnecessary blocks.
The system includes resilience features like circuit breakers, failback chains, and dynamic configuration via Cloudflare Workers, alongside an escape hatch ('break glass') for human override, used in only 0.6% of merge requests.

Hasty Briefsbeta