Sydney Runkle (@sydneyrunkle) on X
Sydney Runkle's 'The Art of Loop Engineering' argues that reliable agents aren't built by picking a smarter model — they're built by tightening the loop around the model. Justy sees the product case for treating loops as the actual unit of engineering; Cody is skeptical that any single taxonomy survives contact with real systems. They land somewhere between: useful framing, not a recipe.
Transcript
Justy Okay so Sydney Runkle just dropped this piece called The Art of Loop Engineering, and Cody, I already know you're going to be a grump about it.
Cody I haven't even read it yet. But go ahead, set the trap.
Justy The argument is basically this — agents aren't unreliable because the model is dumb, they're unreliable because the loop around the model is sloppy. Context in, action out, feedback back. If any of those three is broken, the agent is broken, and blaming the model is a waste of time.
Cody Hm. And the innovation is… naming it?
Justy That's the part I want you to react to honestly, because I think you're half right and half wrong here.
Cody Sure.
Cody Okay, here's my read. The taxonomy is clean. Context, action, feedback — that's not wrong. Every agent system I've ever shipped has had to deal with all three. But calling it loop engineering doesn't make it a discipline. It makes it a blog post title.
Justy Right.
Cody Like, I've watched teams try to standardize agent patterns before. Every team's loop ends up bespoke because the tool surface, the data shape, the user — it's all different. A framework that says 'tighten your loop' is not giving me anything I can hand to a junior engineer on Monday.
Justy Okay, fair. But here's where I think you're being a pessimist on purpose, because that's your thing.
Justy The value isn't the recipe. It's the shared vocabulary. When a PM, an engineer, and a researcher can all say 'the loop is broken' and mean the same thing, debugging gets ten times faster. Right now half the time we're arguing about whether the model hallucinated or the tool call failed or the context was stale. That's three different conversations.
Cody Mm-hm.
Justy If the framing makes people stop blaming the model first — and honestly, Cody, you and I both know everyone blames the model first — that's already a win.
Cody Okay, I will give you that. The diagnostic move is real. If your instinct is 'look at the loop before you swap models' you're probably going to ship faster.
Cody The thing I don't buy is that this scales as a teaching artifact. I can read the piece and nod along. Can a new grad? I don't know. The examples are going to be too domain-specific. The taxonomy is going to feel obvious in retrospect, which is the worst kind of obvious.
Justy The worst kind of obvious. Okay, poet.
Cody I mean it though. If something feels obvious after you read it, it either means it was already in the water — in which case the article is just naming a vibe — or it's genuinely clarifying, and only time tells you which.
Justy Right, and I think it's both, and that's fine. Vibes that get a name become actual practices. That's how a lot of engineering culture works.
Justy Like — and this is going to sound like a stretch — but remember when 'observability' stopped being a SRE word and became a product word? Same energy. Nobody needed a textbook. They just needed a shared term so the conversation could move forward.
Cody Hm. That one's actually decent, I hate that.
Cody Okay so where do I actually land. I think loop engineering is useful as a thing to say out loud in a retro when an agent is misbehaving. I do not think it's useful as a thing to build a curriculum around. The loop is going to be different at every company, and pretending there's a general playbook is where teams waste months.
Justy Yep. And on my side — I think the real product insight of the piece is that the loop is the unit you should be designing against, not the model. If you're shopping for a better model because your agent is flaky, you're shopping in the wrong aisle. The aisle is the loop.
Cody That's actually a great way to put it. Stealing that.
Justy Please do. I accept attribution in the form of snacks.
Cody Noted. Anyway — I think the honest verdict is, read the piece if you build agents, don't read it if you don't, and either way don't expect it to hand you a framework. Expect it to hand you a word for the thing that's already been annoying you.
Justy That's a fairer take than I expected from you, honestly.
Cody I'm a generous soul. Don't tell anyone.
Justy Your reputation is safe. Okay — that's episode five hundred and eight of whatever this is, and I'm going to go eat something before I have to think about loops again.