Hi, everyone--speaker Brendan here. This talk tells the story of my two-person startup with about 120k lines of TypeScript and Python written mostly by Claude where the agent hit a bug that made the size of the code being able to fit into the context window actually drive reconsider semantic compression over being mainstream in language choice.
I found a bunch of other benefits of using Clojure while doing it:
- 68% reduction in the code in scope
- massive reduction in the agent waiting on build times because of image-based applications and the REPL (30s -> 1.4s in a trivial example is shown)
- 4x reduction in type checking mitigated risk of drift between definitions
What went wrong:
- the first pass took 7 days and looked finished, but it was a 7 week tail of deviations from the old system. I had thought the agent could/would use the old system as a functional spec, but I should have taken the explicit step of extracting a written spec first
Caveats:
- some of the evidence for Clojure being better for the agent is self-reporting by Claude itself. Matches my experience, but more hard data would be better.
- the video description says 90k lines became 12k. Those are from my original talk proposal, written before the rewrite had settled. The figures in the talk and on the slides are the correct ones.
Hi, everyone--speaker Brendan here. This talk tells the story of my two-person startup with about 120k lines of TypeScript and Python written mostly by Claude where the agent hit a bug that made the size of the code being able to fit into the context window actually drive reconsider semantic compression over being mainstream in language choice.
I found a bunch of other benefits of using Clojure while doing it: - 68% reduction in the code in scope - massive reduction in the agent waiting on build times because of image-based applications and the REPL (30s -> 1.4s in a trivial example is shown) - 4x reduction in type checking mitigated risk of drift between definitions
What went wrong: - the first pass took 7 days and looked finished, but it was a 7 week tail of deviations from the old system. I had thought the agent could/would use the old system as a functional spec, but I should have taken the explicit step of extracting a written spec first
Caveats: - some of the evidence for Clojure being better for the agent is self-reporting by Claude itself. Matches my experience, but more hard data would be better. - the video description says 90k lines became 12k. Those are from my original talk proposal, written before the rewrite had settled. The figures in the talk and on the slides are the correct ones.
The hooks, skills and style guides we use are at https://github.com/JoinIn-AI/clojure-agent-starter. I'm around today to answer questions, including the sceptical ones.