An AI agent can write Rust; that alone is not a reason to rewrite a working service. Start with a narrower question: where does the time or risk actually come from, and who will maintain the new code?
Compiler feedback helps, but it has limits
Rust can catch many ownership and type errors before a program runs. An agent can use the diagnostic to attempt a repair, creating a useful feedback loop. Compiling does not prove that the program meets its business requirements or is safe in every environment. The Generative Compilation paper studies compiler feedback during generation on specific Rust tasks across seven models; its results should not be treated as a universal guarantee.
What Bun’s migration actually involved
In Bun’s own account, the Zig-to-Rust port was merged after 11 days using parallel agent workflows, an existing test suite and human oversight. The author estimated roughly $165,000 at API prices for the pre-merge model usage. The same account records known regressions that were fixed and continued unsafe code at C/C++ boundaries. Its measured HTTP gains of roughly 3–5% apply to the listed workloads, not every application.
A smaller experiment is often more informative
Profile a representative workload first. If network, database or model response time dominates, reducing local execution time may barely change the user’s wait. Select a small hot path, compare it under the same load and hardware, and test failure cases, security, behavior and rollback alongside speed.
Language choice is a tradeoff: Microsoft’s native TypeScript compiler work uses Go and concurrency. Team knowledge and long-term ownership deserve a place in the decision as much as a microbenchmark.
