Show HN: An experiment in automated long-horizon builds with AMQP 0.9.1

My cofounders and I are in the process of spinning up an AI integration lab, and over the last year we've consumed billions of tokens across multiple American, European, and Chinese models as a force multiplier for building out our products. However, despite our extensive technical expertise, we've still discovered there are long-term ramifications to heavy agentic AI usage. So I decided to run an experiment for us: given a specification, can an LLM fully, concretely, and correctly implement an internally consistent, linearizable distributed system to the same quality as a human?

This repository is the outcome and showcase of that experiment, complete with the entire session history, including tool outputs. To make this project, I only gave the LLM the AMQP 0.9.1 specification and a few requirements: it must use clustered Raft, and it must support up to 9 nodes (light ceiling, but not the final limit). I picked an older but less well-known protocol as it provided a simplistic base with a wide range of features built on top; a messaging platform more than just a broker. While the specification itself is fairly old, it is still very viable and valuable today for a wide range of uses.

For our uses, I chose Switchboard as the experiment as our products are sealed appliances and we need a stable, scalable message broker that's embeddable into our existing apps (embedded distributed systems are easier when it's all the same binary), deployable as a standalone service, and understandable enough by everyone that debugging and diagnosing it is easy. This experiment is not designed to compete with RabbitMQ or other solutions (altho it is a love letter to my former vmware/Pivotal coworkers who worked on RabbitMQ), and I will be maintaining it for our own use cases.

I wrote it in Rust specifically because our code base is written in Rust. While I may talk about the hands-on design, build, and testing of the project as "automated" and "with long horizon capabilities", it is not vibe-coded slop. There are hundreds of cross-platform and cross-architecture tests, and full Jepsen linearization tests on Linux - I am going to put this software into high-compliance production environments and I need it to be stable.

In finality, I hope you find my little experiment in automated, spec-driven, long-horizon builds interesting and insightful. Contributions are welcome, if desired, but with the age of the protocol I don't expect any real consumers other than our lab.

Happy message passing!