Kris Brandow
Works on Go tooling and hosts Fallthrough, a podcast for Go developers that picked up after Go Time ended.
“I have an orchestrating agent and then an implementer and a reviewer. And I found that to work extremely well. The quality of the code I was getting out of the models increased two or threefold. It takes longer, but it's so much better than what I was getting with the one-shot. Even we as humans don't do one-shot code writing. I don't know why we want these models to do it.”
Fallthrough · Hands-On Engineering Podcasts · November 2026
Context
Nick had floated using smaller, cheaper models for programming. Kris describes his own setup, where a cheaper model implements and a stronger one reviews, and says the split improved quality.
Transcript
…month plans, which just don't consume that much usage, I think. Right. Like chat is a very low usage interface for LLMs, especially compared to anything agentic, right? Agentic just has so many more turns that you're just, your hardware is just having to do so much more work than in a chat session.
…use when and for users, every day folks are not going to keep track about that kind of thing. And then you throw API pricing on top and now they have to think about the pricing of each model. And I mean, it's just, that's difficult. And perhaps API pricing, if you need it, you'll know you need it.
Yeah. Yeah. Now, for, and to your point about Sonnet, like I've been using a while ago, I set up my agents.md slash claude.md for all implementation, it uses Sonnet or Terra and it only uses Sol or Opus for like review. And sometimes it Kicks up to like fable for some review tasks if it's like super hard. But like, that's been really good because then, like, the higher agent, the solar, the Opus agent, is just like doing the code review and being like, here's fixes I found. Please do it. I also like, I have like an orchestrating agent and then like an implementer and a reviewer. And I found that to work like extremely well for getting like the quality of the code I was getting out of the models just like increased like two or threefold when I did that. Cause I was like, okay, this is so much better. It takes longer, but it's like so much better than what I was getting before with the one-shot. Because just like trying to do, I mean, even we as humans don't do one-shot code writing. Like, I don't know why we want these models to do it.
No. And it's the more, the more this Euros progresses, the more I realize things like pros and interfaces. And this is the anti-MCP argument, actually, that maybe you should just use CLIs and APIs. The processes that seem to be working are the ones that we had working well for humans anyway. You…
Huh.
Yeah. And I'm reading a bit here from Seeking Alpha, quoting Bloomberg News. It says, some traders had had to use the AI tools to speed up such tasks as monitoring client trading flows, creating written and graphic content, and modeling market changes. One person familiar told Bloomberg. Now…
More from Fallthrough
