← all notes
AI software development LLM Note 011 · October 14, 2024 · 2 min · Updated November 14, 2026

LLM-assisted porting: lessons from Swarm-rb

A retrospective on a Ruby port of Swarm, and why prototype speed does not establish production rewrite costs.

Correction — November 14, 2026: The original article compared this port with an unsupported estimate of OpenAI’s team size and development time, and generalized from it to production rewrite savings. Those comparisons have been removed. The original publication date and URL are preserved.

What I built

In October 2024, I used an LLM to help port the published design of Swarm into Ruby as Swarm-rb. At the time, I reported completing the port in less than 24 hours. That was my account of the experiment, not a controlled productivity benchmark.

Swarm was an educational framework for exploring agent orchestration. Porting an available design benefits from decisions and code that already exist; it is a different task from creating the original framework.

What this experiment does and does not show

The experiment showed me a way to use an LLM while translating an existing design into another language. It did not establish feature parity, production readiness, or a general estimate of the time and money a business could save on a rewrite.

I do not have evidence here for the original claim about how many people built Swarm or how long they took. I also did not publish a measured production migration or a verified customer savings result in this article.

What I would measure in a rewrite

A useful comparison would include behavior preserved, tests and acceptance criteria, integrations, migration and rollback work, operational requirements, and the cost of maintaining the result. Generating an initial implementation is one part of that work.

The question for a team is whether a rewrite improves a specific outcome after those costs are considered. Sometimes the better choice is a smaller change to the existing system.

For my current work and advisory focus, see landongray.me/work and landongray.me/advisory.

The original 2024 draft was written with assistance from Claude 3.5. This revision corrects its unsupported comparisons and separates the experiment from production claims.

I build AI products and help teams put AI to work. These are my notes from the work itself.
Get Landon Notes by email

Keep going

these connect to this one
AI · 2026.03.15 A Non-Engineer's Guide to Building Autonomous AI Agents with Claude Cowork my mom read this one Development · 2025.06.20 The Strategic Software Decision: Code, Train, or Prompt? the decision I get asked most Ruby · 2024.10.01 🍜 Dashi - A Streamlit like Framework for Rubyists [Preview] archived experiment