Connect your Claude, Codex or OpenClaw for full transcripts on millions of podcasts. Two week free trial, cancel anytime.
Start free trial →Whiskey Web and Whatnot · Hands-On Engineering Podcasts · November 2026
The hosts were comparing coding agent tools. Adam Argyle says he tested a model-agnostic harness against the one the model's maker ships, and the native one won on both cost and output.
…But Shopify did do it. And I can't say exactly what, but Shopify did it again this past seven days, took a humongous piece of corporate paid enterprise software and said, not using you anymore. We built it ourselves. And that would be about the fourth time I've seen him do that in the past year.
…you want, use the best frontier models. Like, you're still kind of a startup-y vibe, like you can move fast. Like, we have a lot more red tape at IBM and stuff that's been happening. But we have cool tools back now. Like, I can use Astra through Copilot with Open Code. So that's what I'll be doing.
Wait, I want to, I want to premise something. My Pi usage of Claude and Astra. So my Pi usage of Opus and Astra is poorer and more expensive than if you just use the Codex CLI. I haven't used Open Code as much, but I'm of the current. Opinion that the harnesses built for the models are superior to the agnostic hardest harnesses that allow multiple models. I went back to Pi because there's like a Pi craze going on with Toby and other folks at Shopify. And I was like, what am I missing? I've used Pi for a long time. I was like, I was one of the first people with Pi. I love Pi. And then I was like, I don't need this anymore. It's kind of like getting crappy. I'm noticing that Claude is better. And I went back to Pi and I doubled my cost and did not increase my productivity at all. However, using Codex CLI, specifically with Astra, is superior to Pi using Astra. And I think that is easily explainable. And so I just want to, I just want to usher you to do that. Yeah.
It's like how Apple builds hardware and software. It's like that's how they work well. They know the strengths
and weaknesses of their models. And so they'll play into those. Yeah.
Yeah. So it's always going to work better, even if it's like, they might even be explicitly just telling it, like, hey, if this is coming from our stuff, just work better. Like that could be built into like the core directive of like how they work. And we'll never know. So