- AI coding with Rust helped one engineer write over 100K lines of production code in roughly four weeks.
- AI coding with Rust reached 300K operations per second — up from 23K — through focused performance optimization.
- Code contracts written by AI and verified at runtime caught a critical Paxos safety bug before it reached production.
- Lightweight spec-driven development, using single user stories as the unit of work, proved more effective than rigid documentation.
One Engineer, Three Months, 130K Lines
AI coding with Rust has produced some eyebrow-raising productivity claims over the past year, but few come with this level of specificity. Zifei Huang, a distributed systems engineer, spent roughly three months building a modern replacement for Azure’s Replicated State Library — the multi-Paxos consensus engine that underpins a significant chunk of Microsoft’s cloud infrastructure. The project produced more than 130,000 lines of Rust. About 100,000 of those were written in four weeks. Whether you find that inspiring or alarming probably depends on what you think software engineering actually is.
The original RSL is over a decade old. It works — Microsoft has run it at enormous scale — but it wasn’t designed for the hardware landscape of 2025. No pipelining means new requests queue up while a vote is in flight, bloating latency. There’s no support for non-volatile memory, which is now standard in Azure datacenters and can dramatically cut commit times. And it predates the widespread adoption of RDMA networking, which is now everywhere in Azure’s fabric. Huang’s goal was to start fresh, fix all three gaps, and do it fast — using AI agents as the primary development engine.
The AI Coding With Rust Workflow That Actually Worked
Huang didn’t settle on one tool. He experimented with GitHub Copilot, Claude Code, OpenAI’s Codex CLI, Augment Code, Kiro, and Trae before landing on a two-agent setup: Claude Code and Codex CLI running from the terminal, with VS Code handling diffs and smaller edits. The CLI-first approach, he says, creates an asynchronous flow that keeps him out of the loop on mechanical work and focused on decisions that actually require judgment.
There’s a psychological dimension here worth taking seriously. Huang pays $100 per month for Anthropic’s Max plan. His framing: if he doesn’t kick off a coding task with Claude before bed, he feels like he’s wasting money. It sounds trivial, but that kind of forcing function — turning a subscription cost into a daily commitment — is a real behavioral trick. When Codex CLI launched and rate limits became a bottleneck, he added a second ChatGPT Plus subscription and split the week between them: one account for Monday through Wednesday, the other for Thursday through Sunday. It’s the kind of scrappy optimization you’d expect from someone who used to hand-tune memory allocators.
Code Contracts: The Correctness Layer AI Actually Gets Right
The obvious question when you hear
