| title: | Why we throw out our AI agent and rebuild from scratch every 6 months |
|---|---|
| description: | Life moves pretty fast |
| pubDate: | 2026-02-03T00:00:00.000Z |
| author: | Steve Krouse |
Over the past three years at Val Town we've been copying the best AI coding tools (and occasionally getting copied ourselves!). We've done five throw-away-the-code-and-start-from-scratch rewrites:
- v1: ChatGPT-inspired
- v2: ChatGPT-inspired, with tools
- v3: Claude Artifacts-inspired
- v4: Open source, in "user space"
- v5: Claude Code-inspired
- v6? More on that in a minute
Some of those versions felt like the bell of the ball while others were just ok. The through line is that we feel good about investing in timeless developer principles and tools—the AI will find its place. Five rewrites in, this is the story of why it's all worthwhile.
The pace of AI is so amazingly (or fatiguingly) fast that it's proven easier to throw out all our code every six months. AI is changing devtools along two dimensions:
- LLMs keep getting better: smarter, faster, cheaper
- We as an industry keep finding better form factors: tab completions, tool use, agent loops
It's hard to predict how programmers will want to code a month from now. The pace of change makes it incredibly hard to invest deeply in any one approach.
- LLMs are improving on (1) price, (2) performance, (3) cost like ~100-1000x every 18 months. That's faster than Moore's Law, which itself was amazing!
- On top of that^, the industry is iterating on AI interfaces very quickly. The way programmers interface with code is changing super fast
Before building Townie v5, our intention was to thoughtfully craft a good MCP server so that you can stay where you are (in your favorite AI coding tool). But MCP clients just aren't good enough yet, so we built our v5 in-browser coding bot on top of our MCP server. I would love for the AI labs to realize that in addition to Consumer and API companies they are also Platform companies. I want them to take this responsibility and opportunity seriously and invest in a proper "App Store." We could focus on building the best coding platform while the AI providers handle the AI things — billing, rate limiting, combating abuse, etc.
Before founding Val Town in July 2022, I ran the Future of Coding podcast and community. Boy, were we completely blindsided by AI. We didn't predict or invent the future of coding at all!
But I'm not an AI hater. I'm actually quite into it. My fiancé jealously calls ChatGPT my "girlfriend." If we were fundraising right now, I might feel short-termist pressure to lie through my teeth telling you how Val Town "was founded to reimagine what computing could be with AI." But I don't have to! Because the whole point of AI devtools are to multiply your powers as the programmer working on the code. And we built Val Town for you the programmer and your code. AI just fits in naturally.
Jeff Bezos said he could afford putting lots of effort into things that won't change in 10 years. He knew Amazon customers would always like lower prices and faster shipping, just like we know that you will always like 100ms or faster deployments and lower infra setup costs.
So will we throw away the code and rewrite Townie again in 6 months? Yes, probably. And it will continue in our tradition of fitting together our favorite AI coding patterns of the moment with Val Town's timeless bets.
Thanks to my "speechwriter" Pete for helping me write this.