An agent sends a payment request. The provider accepts it, the connection drops before the response arrives, the worker times out, the orchestrator retries. The customer is charged twice, and a Rust rewrite would not have prevented it.
Rust's ownership model removes use-after-free bugs and data races in safe code. It can't know whether a remote API finished an operation before the network failed. That takes durable state, idempotency and a recovery protocol that separates safe retries (a read-only search) from operations that need reconciliation (a payment).
Raw speed is a weak argument too. In an illustrative run of 4 model calls at about 800 ms, 6 tool calls at about 150 ms and 100 ms of orchestration, removing orchestration completely saves about 2.4%.
Where Rust earns its place is concurrency: how many agent sessions one worker sustains in a fixed memory budget at acceptable tail latency. Memory safety, concurrency correctness, distributed execution correctness. Rust covers the first, helps with the second, and leaves the third to your architecture.
Full piece, with the state store, queue, worker and recovery design: Why Rust might be the most important language for the next era of AI
Top comments (0)