What should count as proof when AI writes a software migration in a language the team does not know?
Anthropic’s platform team reportedly had no Rust developers, used Claude to write a Rust port of its Messages API, and verified it against live shadow traffic, cutting CPU use about 40–50%.
The useful detail is the verification approach. Shadow traffic means using copies of live requests to check a replacement alongside the existing system. That can help test whether generated code behaves as expected on real requests; it does not, by itself, establish that every important case is covered.
Before attempting a similar migration, engineering leaders should define:
- Which outputs and failure behaviours must match.
- How differences will be detected and reviewed.
- What conditions make the resource comparison meaningful.
- Who can maintain the code and approve a switch.
The practical lesson: make acceptance criteria and maintenance ownership explicit before treating an AI-written port as ready to run.