Ronan: Boris Cherny, the fella who created Claude Code? He deleted his whole IDE back in November. Just gone. Hasn't opened it since.
Tara: Deleted it? His entire IDE? That's a... a drastic step, isn't it? What's he doing instead, if he's not coding?
Ronan: Ah, here's the kicker. In the last, what, thirty days? Pretty much all his contributions to his own project, like, ninety-something percent, were written by an AI. And it's an AI he manages, mind. Not him, the machine.
Tara: Oh, that's... that's a bit unsettling, actually. So he's just, like, overseeing the AI? He's the manager for a robot coder?
Tara: So, this is what everyone's been on about, isn't it? This whole 'loop' thing. Peter Steinberger's tweet, it blew up, didn't it? Millions of views. He said you shouldn't be prompting agents anymore.
Ronan: Yeah, that's the fella. 'Design loops that prompt your agents,' he said. And the comments, oh my god. Everyone was arguing, and one person just, like, summed it up perfectly: 'nobody knows but him and Boris.'
Tara: That's it! It felt like that. Everyone talking, but nobody actually getting it. So, alright, what is a loop, then? Is it just... a script? A fancy name for, like, a cron job?
Ronan: See, that's what I was thinkin' at first, right? A `while` statement, somethin' simple. But no, the article explains it's a proper program, something you write yourself. Not just a little bit of code.
Tara: So, wait, it's a program... that tells the AI what to do?
Ronan: Yeah, that's it. It prompts the AI.
Tara: And then the program, not me, it reads what the AI made? And decides if it's good enough?
Ronan: Exactly, it's checking its own work, or rather, the AI's work.
Tara: And if not, it just... prompts it again? On its own?
Ronan: That's the key, yeah. You're not typing the prompts. You're writing the thing that types the prompts.
Tara: Right. So, the AI is like... a part of my program. A function, maybe?
Ronan: A subroutine, yeah, that's it. In this bigger automated system you've built. It's your job, but it's moved up a level. Instead of doin' the coding, you're orchestrating the whole shebang.
Tara: Hmm. You know, this feels a bit like... like when Infrastructure as Code came in, right? We used to SSH into servers, manually type commands, configure everything by hand. And then suddenly, you're writing Terraform, Pulumi files. My job didn't vanish, it just... shifted. You're not doing the 'thing,' you're writing the 'recipe for the thing.' That's it! The recipe.
Ronan: Ah, 'the recipe,' I like that. Spot on. It's not gone, then, just... completely different. And the article actually has a good bit on that, gives a bit of a history, a 'ladder' for these loops. Helps clear up a lot of the confusion.
Tara: Yeah, it breaks it down, like, five stages. The first one, 2022? Just this academic 'while-loop' from the ReAct paper.
Ronan: Right. Human supervising.
Tara: The model, it'd reason, then call a tool, read the result, then just... repeat. But always with a human watching it.
Ronan: A human supervision job, yeah. And then 2023, that was AutoGPT. Now that was a bit of a disaster, wasn't it?
Tara: Oh, it was! I remember. You'd set a goal, let it go... and it would just spin forever, wouldn't it? Doing absolutely nothing useful. Really gave agents a bad name.
Ronan: Yeah, completely. Then 2025 rolled around, and you got the 'ralph' loop. Just a bash one-liner. Dead simple, like.
Tara: A bash one-liner? That's it?
Ronan: That's it!
Tara: And that was a whole stage?
Ronan: Yeah! But the clever bit was it reset the context every time. Kept it from getting bogged down in all the old conversations.
Tara: So that took us to Stage four. Where it got, well, productized. Tools like Codex and Claude Code, they shipped a '/goal' command. Spring 2026, I think it was.
Ronan: Right. So it's still running the 'ralph' loop, but it's got an AI check to say, 'yeah, this is done'?
Tara: Exactly! A smaller AI model, confirming the task. So it's not just running forever.
Ronan: Ah, see? That's gettin' a bit smarter, then. And the fifth stage, that's where we are now, isn't it? What Steinberger and Boris are actually talkin' about?
Tara: Yes! The real deal. Multi-agent orchestration.
Ronan: Multi-agent orchestration?
Tara: Loops supervising loops. Running at the same time. On a schedule.
Ronan: Like, just... everywhere?
Tara: Yeah, everywhere. And they've got this durable, git-backed state. So if the server dies, they remember what they were doing. It's a proper system, not just a little script.
Ronan: Okay, that's a lot more advanced than I was thinkin'. So, what does Boris say? How do you actually run these things, then?
Tara: Alright, he's got five main tips. First, 'auto mode.' The agent can just act. No constant asking for approval from you.
Ronan: So it's not like, 'Can I do this? Can I do that?'
Tara: Exactly. Then, 'dynamic workflows.' Your main agent, it orchestrates hundreds, maybe thousands, of smaller agents for big tasks.
Ronan: Right, so it's not just one agent, it's an absolute army of 'em.
Tara: An army, yes! The third one: use '/goal' or '/loop' commands. That's to nudge the agent, keep it going until the task is truly, properly finished.
Ronan: Right, keep it on track.
Tara: And number four: run it in the cloud. That way, you shut your laptop, go home, the loop just keeps running.
Ronan: Which is grand if you want to, you know, actually go home for the evening. Or go to the pub, maybe.
Tara: Yes! Gives you your evenings back. And his fifth, the one he really hammers home, 'self-verification.'
Ronan: Self-verification.
Tara: The loop, it has to check its own work. End-to-end. It needs to know what it's done, if it's correct. Otherwise, what's the point?
Ronan: See, that last one makes perfect sense. Otherwise, you've just got an AI confidently makin' absolute gobshite mistakes, haven't you? But... there's somethin' I'm still caught on, Tara.
Tara: What is it, Ronan?
Ronan: All this talk of scheduled tasks, runnin' in the background... it's just a fancy new name for cron jobs, isn't it? Like, we've had things running on a timer forever. Cron was invented in, what, 1975?
Tara: You know, that's a good point. The article actually does address it. And you're not wrong, the scheduling part is similar. Boris actually uses cron for his loops.
Ronan: He does?
Tara: Yes! But the real difference, the key, is what's inside that cron job.
Ronan: Go on, then.
Tara: A traditional cron job? It runs a fixed script. It does the exact same thing, every single time.
Ronan: Predictable.
Tara: Predictable. But this loop? It runs a model. That model looks at the current state, decides what to do next, does it, checks if it worked... and then decides if it needs to keep going.
Ronan: So it's a decision-maker, then. Yeah. But it's still within the parameters I set, right? It's not just gone off on its own little adventure.
Tara: No, it's not fully autonomous in that sense. But it's making dynamic choices based on feedback, its current state. That's well beyond a static script. And when you start stacking these loops, letting them supervise each other, giving them durable state... cron alone can't really express all of that. That's where the—
Ronan: So it's the intelligence inside that's the difference. Not just the schedule.
Ronan: Right, I get your point about the decision-making. Still, I can't help but feel a bit of... a sinking feeling, you know? Just... another new thing to learn. I was just getting my head around prompt engineering and now, what, it's all changed again? The goalposts... they just keep moving.
Tara: It's that treadmill, isn't it? Always something new. But, I think this time, it's actually... manageable. The article, it gives a really clear way to start. It's not some big, you know, mysterious thing.
Ronan: Oh? Is it?
Tara: Yes! Boris Cherny himself, he gives you the actual command. Right there, in Claude Code.
Ronan: A command? To start?
Tara: Yeah, just: '/loop babysit all my PRs. Auto-fix build issues, and when comments come in, use a worktree agent to fix them.'
Ronan: That's... that's very specific, actually.
Tara: It's practical. It's actionable. It's not about being left behind; it's about changing how you do the work. Not what you work on, just... the how.
Ronan: So, no more scratchin' your head over it, then? It's not some mystical beast, these 'loops.' Just a smarter way to get the AI workin' for you. Not the other way 'round.
Tara: It gives you a way to engage with the tech, instead of feeling overwhelmed.
Ronan: And they even give you the exact command to start. That's, uh... that's good. I'm Ronan.
Tara: And I'm Tara. This has been Manish Chiniwalar's Station.