One developer built a multi-Paxos consensus engine in Rust over 3 months using AI agents, scaling throughput from 23K to 300K ops/sec.
Key Takeaways
130K lines of Rust covering full Azure RSL feature set (multi-Paxos, leader election, snapshotting, config changes) written in ~6 weeks via Claude Code and Codex CLI.
AI-generated code contracts (preconditions, postconditions, invariants) caught a subtle Paxos safety violation before production; GPT-5 High outperformed Opus 4.1 for contract quality.
Property-based tests generated from contracts explore randomized inputs automatically; 1,300+ tests now cover 65%+ of the codebase.
Performance loop: AI instruments latency, writes Python quantile scripts, proposes optimizations, repeat; gains came from zero-copy, removing async overhead, and eliminating lock contention.
Lightweight SDD replaced rigid spec docs: /specify generates user story specs, /clarify triggers AI self-critique, then single-user-story plan mode drives each coding session.
Hacker News Comment Review
The sole commenter argues AI agents likely produce bloated codebases and that raw line count is a poor metric for productivity or quality.
Notable Comments
@staszewski: “I don’t know whats impressive in lines of code generated by agent” – flags that agent output volume signals waste, not achievement.
