“Sometimes you talk about taking a 15-minute run down to a five-minute run. And maybe on the surface, that doesn't sound that impactful, but that's like four iterations per hour versus 12 iterations per hour, which is a pretty meaningful difference.”
sputnikintelligence.com
A co-host was excited about AI agents running CI themselves so pull requests arrive already green. Manges, who builds a CI platform, explains why faster feedback loops matter by counting iterations per hour.
…a sandbox using RWX Sandbox Exec and make sure the full CI pipelines are passing, you know, with RWX run. And then that's it. By the time a human's in the loop, by the time somebody's having to review the change, they've been confident that the agent has already verified that everything is working.
…through GitHub and I'm like, okay, well, this one failed and I need to go tell that agent, like you failed, like try again. So the fact that it could just kind of run the CI itself and confirm. And then when you get to GitHub, it's already green, ready to go. That would save me so much time. Jesus.
Faster feedback loops are really meaningful. And, you know, sometimes you talk about taking a 15-minute run down to a five-minute run. And maybe on the surface, that doesn't sound that impactful, but that's like four iterations per hour versus 12 iterations per hour, which is a pretty meaningful difference. And yeah, I do appreciate DHH's, I think, ability to innovate and be contrarian, find new ways of doing things. And I think he's certainly right that, you know, now local chips are so fast that they're faster than server-based infrastructure. And for some apps, just running the entire test suite locally probably does make sense. But I'm sure some of you guys, a lot of people listening, you know…
Where do you kind of see like the future of CI CD going now that we're all in on agents? I mean, especially if you want to talk about like, there's been a lot of security stuff happening this week. And I feel like now more than ever. We went and started auditing a bunch of stuff where we're like…
Yeah, for sure. And I mean, you guys opened the show talking about some of those security vulnerabilities. And I think this is just going to be the increasing reality of the world that we're in now, especially with AI's ability to surface some of those vulnerabilities. And, you know, I wasn't…
Yeah. I use Claude and Codex together. Come on, Chris, get your game up. I'm lucky to work at a place that's like, yeah, if you want to experiment with that, go ahead. Here's XYZ, whatever you need to try it with GitHub reliability lately. And we switched to self-hosted runners, but the setup for…