The Great Loops Debate — Dex Horthy, Geoff Huntley, Ian Livingstone, Greg Pstrucha, @insecure-agents

11,498 words · about 46 min to read

Debating the Delta Between Loop Hype and Reality

0:12 Hi, welcome to the great loops debate. My name is Ali Howard, the host of the insecure agents podcast. I also am a member of technical staff at Key Card and I'm super excited to bring the loops debate to you here today. You might recognize some familiar faces on stage today that did the MCP debate with us at AIE code back in November. Um, that was so much fun. We thought maybe we would do it a second time. Um, but this time we would debate loops and instead of just Ian versus Jax, we uh, we recruited some more people for each of them to have on their team. Um, so here on stage with me today to debate loops, we have Ian Livingston, CEO and co-founder of Key Card. We've got Jeffrey Huntley, the creator of the Ralph loop. Um, Greg Kostruba, developer at Century. And we've got Dax Raad, who you all know, um, CEO of Human Layer. Um, so super excited to have all of these people here with us today that are very close to loops engineering and also um, software factories to debate this topic.

1:08 Um, so you might be wondering what are we actually debating today? Like aren't loops in and engineering um, in software factories pretty well uh, promoted and understood and we all kind of agree that's like where the industry is headed. Like what is there to debate? Um, it's a great question. What we're here to debate here today, the core thesis is there is or is not a delta between the hype behind loops and what actually works in practice. Um, and we also can debate, you know, is now really truly the largest inflection point we've seen towards fully autonomous software factories. Um, so super excited to cover in the debate today uh, loop history. We'll debate that.

1:44 We'll debate the loop anatomy, like what makes a good loop. We'll also debate the future of the loop, um, future of loop engineering and what we need to see from that in order to make software factories truly something that every organization is able to build. And today's um, is going to be a Oxford debate format. So we're adding a little more structure to the debate this time this round where every single person will have a timed window to give their response. And so we won't have people running over.

2:12 And the way this works in practice to judge the winner, which we'll judge in real time, is the winner will be the team that changes the most minds. So I want everyone in the audience right now to take a minute to think about whose side you're on. Lock that in, remember that. And at the end I will ask you to raise your hand and say, "Okay, actually I was changed my mind. Actually I'm now team Dexter. I changed my mind. I'm actually now team Ian."

2:38 So to present the two sides to help you make that decisions, I'll talk about the first team, which is Ian and Jeff's team. Their team no Delta. Loops are absolutely worth the the hype today. People can get up and running with them easily, build them. They're an important step up the autonomy curve and towards real software factories. Key points to look out for for team Ian and Jeff is that loops are core unit of engineering. With the right discipline, infra, and tests, loops are highly effective. And the best practices for those have emerged.

3:12 For team Dex and Greg, their side believes there is a delta between the hype behind loops and what actually works in practice. The way we are doing loops today is wrong. Loops are not a silver bullet and there is no magic. Key points to look out for. The hype is out running the discipline and a software factory can run the mechanical spec'd out test covered slices unattended. It cannot on and honestly decide whether it built the right thing. So you still need engineers in the loop essentially.

3:40 So now that you know a little bit about what what both sides are about, take a minute to internalize that. Think about, okay, like where do I stand? Am I team Dex or am I team Ian? And then end we'll ask you again. So, from here we'll go into the beginning of our debate. Um each member

Loops Represent an Inevitable Shift in Software Architecture

3:57 of each team is going to give a 4-minute monologue on why they're defending their side. Um and just start us off, we'll have Jeff. Jeff, I'll give you 4 minutes on the floor. Why are you Why are you pro loops today?

4:10 It's because it's somewhat inevitable. Um I I have first uh basically if we wind back time 2 and 1/2 years ago when I was a tech lead over at Canva and I was seeing all the engineers just prompting and prompting and prompting and they they they they were in the loop. And I'm like, "Wait a sec. This could be programmed. This is a programmable thing." And it just became a really inevitable. Um loops are uh somewhat uh whilst Ralph might be a bit of a meme and cetera, there was actually deep thought into it.

4:40 Um it's essentially applying this as if this is a new form of CPU architecture and figuring out the behaviors of this and how to do it. And through that I was able to reduce it down to a bash loop. Now, it is not a complete silver bullet, folks. I have my deepest concerns next this time next year at the conference. We're going to see a whole bunch of talks saying how factories fail and how loops fail. Um these are things These are things that we are still yet to figure out.

5:13 Do you remember the early days of Kubernetes where everyone was just doing Kubernetes? So, it's here. It's inevitable. It it It is here to stay. Like programming the machine and automating your job function is the expectation of employers. Categorically. And um I don't see myself going back to writing code by hand. It's been 2 and 1/2 years since I manually wrote code by hand.

5:40 Um I autonomously factor code from one code base to another code base.

5:45 [snorts]

5:45 I find something in GoLang and I'm like, "Oh, I'm in TypeScript." So, I I I run a loop, and I just autonomously ported across. And like, even for product managers and product manager research, it's easy for us to index on what it means as software engineers. We can focus just on software engineering, but think about something like, you want to do product management research on all the linear tickets. Well, there is a termination. It's actually defined when you've enumerated all of those linear tickets.

6:14 So, that's easy. So, there's a lot of nuance in here. But like, one If you ever done any product management research, and you started running these loops, and be able to like compress time and the amount of time to do that research, it it's inevitable. We got this new programmable substrate. We got to figure out how to use it, where it's good to use, and um I know I'm meant to debate like that it is the thing in all thing, but in all things, there is no silver bullet.

6:43 And we're going to figure out how we're

The Hype Outruns Discipline in Fully Autonomous Loops

6:45 going to be using it effectively over the next year.

6:51 Thank you. All right, next we'll have Dex. Tell us why there's a difference between the hype that exists between Loops today and Loops themselves.

7:00 Okay, cool. And um what are the What are the rules around personal attack in this debate? Is this encouraged or

7:06 You can do whatever you want with your 4-minute monologue. That's That's the rules, yeah.

7:10 Um It's funny, this is reminding me of the debate last year, where we're we're all kind of arguing, and uh maybe at the end we all I'm going to be convinced that Jeff is right, but Jeff is going to be convinced that I'm right, and maybe we switch sides by the end. Um

7:26 [clears throat]

7:27 Yeah, I think uh the the the basic take here is is not whether Loops are good or bad. I think um it's funny you bring up Kubernetes, because Kubernetes was this thing that took us seven or eight years to get right.

7:40 Yep.

7:40 Uh and before that it was cloud infrastructure, and you could argue that like it took us seven or eight years of cloud infrastructure to get to Kubernetes to then seven or eight years to get that where it was really usable by everybody. Uh, and Kubernetes is actually built on loops. It's built on control loops, but they're deterministic loops. And we've actually figured out exactly what types of things that, uh, small isolated tasks that can be sort of owned by one system. I think this is actually the biggest value in loops is that you can pick a small, sort of desired end state and feed in the current state of the world and have an agent or a deterministic system kind of progress towards that desired end state.

8:24 Um, the challenge I have with the hype is we were already in a world where it where the the the prevailing mantra was see if you can get to a point where you don't have to read the code anymore. And, uh, even before loops, it was like just prompt, just go. And this idea of I don't even prompt anymore, I'm even at a level higher up implies that I am like taking even more of a backseat to the architecture of my code. And I think my biggest like point in terms of like the hype is outrunning the discipline is that we are all looking for magic, you're all looking for a silver bullet, we're all looking for something that will take away that horrible part of our jobs that we all hate, which is like reviewing code. Uh, some people enjoy it. Really good pull requests are fun.

9:14 But that we can somehow prompt our way out of, um, this this challenge that models have of like, okay, the code is pretty If you've ever reviewed a fully, uh, lights off sort of, uh, no one read the code before they sent it to you PR, I'm sure you've had this, uh, probably not great experience. And I think I've seen lots of people try to apply AI to this problem of hey, we have review bots and we have all these things.

9:42 But it it doesn't seem it doesn't feel to me like it's working. I haven't seen proof uh in any any of the discourse publicly or any of sort of more private conversations we've had with people trying to put this into practice that we are we are at a point where we can just kind of like step up an abstraction level. I actually think we need to step down an abstraction level if anything. Um so, I think loops are there are good things about loops and we should be doing them, but the hype is is making us

Software Development Has Always Been Loop-Driven

10:12 feel like there's a magic answer to this and it requires a lot more thought and care than the uh the the Twitter sphere would have you believe.

10:22 Excellent. Yeah, good points. All right, I'll kick it back over to you Ian for the pro loop side.

10:28 Absolutely. So, I think first and foremost, I'm coming for you Dex. But in reality, um let let's let's take a step back. And let's talk about like what is software engineering in the first place and also just like remove the word engineering and talk about like development. In inherent in building a system, whether it was 50 years ago or it was a thousand years ago or it's today, it is a is a loop is at the core of I try something, I learn something, I apply something.

10:54 And all we're really talking about is how quickly we can expedite that process, right? And really what we're doing is removing what used to be human judgment in that process. So, the speed at which it is to generate something and removing that from humans typing. And you know, tab completion is a version of of auto complete and what is this stuff other than like really like a much better version of tab complete. Except now we give a much higher level version of what intent instead of just typing tab, right? And so, I think the premise is loops are at the core of everything we build already. They were at the core of how we built software 30 years ago.

11:28 What is CICD, PR pull requests, design review, feedback from customers other than just driving a loop. And the question is, how much of that process can of the which is deeply subjective and requires reasoning, can we move out of the human brain and into the brain into these like non-deterministic models? And the underlying question for all of us is more about verifiability. Software is one of the most unverifiable things in the world because ultimately at the end of the day, it is most things can become true or false one way or the other. And my premise and my my point I want to make here truly is as humans interact less with software, which is how does a human interpret what that software is doing, how does a human interact with that software, how does a human use judgment to navigate that software, the subjectivity of what that software needs to be through the human-computer interface reduces and becomes a much more verifiable problem because it becomes more constricted to specific APIs. And so over time, it's both the fact that at the core of software development is loop-driven anyways, what is lint than a

12:28 feedback loop, and that creates a verifiable thing, and as more humans are less interacting with software, you have less UI and UX and less subjectivity and what and how we interpret those things, you will become much more loop-driven and you'll become much more verifiable in a way that wasn't previously not possible.

Token Economics and Quality Limits Stifle Infinite Loops

12:45 Awesome. Yeah, good points for sure. All right, Greg, you want to close this out with the anti-loops or the there's a delta between the type.

12:53 Absolutely. I do think that the way that I would start this is um there is a lot of hype. There is a lot of FOMO. There is a lot of I'm looking at Twitter. I'm seeing what people are talking about. Am I missing out? Am I doing something wrong? Am I holding AI wrong? Uh should I be catching up? And that is um very stressful. And I think it boils down to two points for me. One of them is when you are generating code with AI in any manner, loops or no loops, are you happy with the output? Do you think the output qualitatively is what you need it to be to do whatever you are trying to do to get to the desired state. Um and if so, I would love to learn from you. I have not get there. Uh I think that the best way that we are improving that is both with model intelligence, but also as everybody here seems to agree, semantic verification as much as we can do statically, we should do statically.

13:48 That's that's sort of one big thing where I lean on and in practice end up reading uh code that is sort of post semantic verification and it's still crap. I still have to do a lot of iteration on that uh on my own and still have to steer it towards the right architecture, tell it where it should be simplified. Um And so that's one big step. And there is a lot of things you can help yourself with by throwing more tokens at the problem, but one of the things that the current sort of hype-based discourse uh leads you to believe is that you can just have loops on top of loops on top of loops and orchestrate that or orchestrate your problems of quality away by more tokens.

14:29 And that brings me to my second point, which is the economic viability of of the way that we are using agents today. And I don't believe that this is sustainable. I don't believe that like you when you are at the company, especially a larger company, you have to ask yourself what is a good budget for an engineer? Is it 10K a month? 100K? 1 million dollar a month for for a token spend? At some point that that just starts cracking and it's not sustainable in the way that we are doing it today. That said, I am also writing code with agents and I also use some loops um for some specific flows. It just depends. There is nuance. You know, if you go to Twitter, Twitter has no nuance. Um but there is actual nuance to the conversation and there are specific tasks and and jobs that uh you can already loop on and be and be getting

Engineering Guardrails Keep Non-Deterministic Models on Track

15:18 pretty reasonable results.

15:22 Thank you. All right, Now that we've heard from each of the debaters on stage more about their stance and their opinion, we'll move into the main debate portion. The first section of our debate is going to focus on the history of the loop and why now is a major inflection point or not for loops engineering and also software factories. I know people will have said that you know, vision models have improved a lot and so they'll be able to verify work the agents have done was not possible previously. Context windows have improved, therefore memory is improved and we can now track work in a loop that maybe we couldn't have before. So now we look at this like loops history, where did the loop start from? Was it you know, Jeff Huntley's like Ralph loop?

16:05 We'll get into some of those questions and debate here. It'll be the questions we targeted now at a very specific like single person and only they will get to respond and they'll have 2 minutes and 30 seconds to respond. So our first question is Anthropic took Jeff's concept of the Ralph loop, absorbed it into their platform and created a series of three commands, loop, batch and goal. The goal command is designed to keep going until a condition is true. Agents are very determined and the whole point of this command is to keep going and iterating through ways to solve a task until it's done.

16:40 Ian is a security expert in the room. How are you confident agents can stay aligned to their task and not overstep their intended permissions while ruthlessly pursuing their goals?

16:50 I mean I think if there's any evidence, I'm totally not convinced that that's possible. In fact, I think what we've seen is as we scale with these models and as we use reinforcement learning, they're inherently incredibly goal-seeking. And so we're now we're seeing them finding exploits and vulnerabilities and escapes that you know, humans through hundreds and thousands and thousands and thousands of hours and attempts and attacks have never been able to find. So I I don't think inherently the model itself in any capacity can keep itself aligned Um to aligned and safe, right? And safe is a word that I don't love to use because it implies a bunch of things.

17:27 So I'm I'm not I'm not I don't have good belief that the model itself can actually do that. I don't think it can reason. I also don't believe holistically that a model can tell good from bad and I can't tell whether it's doing something malicious or unaligned. It is not alive and it doesn't actually self if it doesn't have air. It doesn't deal with the fact that hey, if I do something wrong, no one's going to love me or or want to be my friend. And if I do something good that someone's going to praise me, it may seem that way, but it is just a probability distribution at scale. So I'll still I'll still be your friend even if you I know Dex will hug me after this even though it seems like it's not true. So broadly speaking, I don't think that comes from the model and it doesn't come from the loop. It's about the infrastructure you build around it and how you enable that infrastructure to actually enable you to take advantage of these loops. And as the models get better and as the underlying infrastructure and platform we build to enable these these feedback loops and loop automation and software development, whatever word software factor whatever word word we want to use for this conjunction of stuff that sits

18:28 on top of a probability distribution, we will be able to have better guarantees, but we certainly are not going to be able to I do not fundamentally believe or ever believe that some type of alignment or reinforcement learning is going to result in a model ever being 100% safe in any capacity. And if there's any evidence, it's that as these models get better, the most important thing to remember is they actually become higher goal seeking and higher capable in terms of finding exploits to achieve their ultimate goal.

18:55 No, I would concur with that completely. Um if you the the most concrete thing you can do to secure your environment is just not have secrets as files.

19:03 Yep.

19:04 Um if you ever seen the behavior where it wants to deploy a web service or what else have you and the token's not privileged enough, it'll start goal seeking on the file system looking for higher privileged tokens credentials. You do not want to get in the way of an agent wanting to do its a goal.

19:21 Okay, so next question is directed to Jeff. Jeff, your original post from last year said Ralph was best for greenfield work. Today it seems engineers are running loops on existing code bases to improve latency, eval's, or refactor parts of their back-end code. What's changed that suddenly makes loops more broadly usable today?

19:42 Um it doesn't matter how good models get, folks. The models have been good enough for at least the last year. Um what has changed is people's understanding of that. So society is only able to adjust at a rate. For example, I hypothesize it's it's Christmas breaks. Because the models back in November were released in November last year. In December, there was no real new models.

20:12 What the difference was people had time. They could actually sit down and play with it. They they had the realization that these have actually gotten really good. So the reason why is it because it works,

1 Highlight · Running Loops Economically Beats Mass-Market Hiring

20:24 folks. These LLMs generate code better than you can hire you can actually hire for. If you think in the broad mass of software developers or coders, these LLMs generate code better than any software developer in the mass market that most founders can actually hire for. It it's it's sad but true. Um now why loops? It's really simple.

20:50 Cuz if you run it in a loop, it works out to $10.42 an hour. Calculation index did back about a year ago now.

20:59 Yeah, August we did the we did the hackathon where we we we copied all of

21:03 the sponsor tools. We rewrote a bunch of Python libraries in TypeScript.

21:07 Yeah. So, like concretely, like loops. I've come across so many engineering managers and founders and they've got these complex tech stack. They're running on four or five different programming languages, etc. And they run a loop and they've got good tests on these type of tech stack. And all of a sudden their complexity is they're just managing one tech stack. It works. Go to YouTube. You've got all these software developers now who are now software developers because it's software development as a profession has been commoditized.

21:34 Some deep thinking to be there. And they're just on YouTube and they're like, "Yo, check out rough loop because it's a I went to sleep and I woke up and it works." Like boss, it's meme-y, it's punchy, it works. But there are there are problems with folks. Like I originally described it to should be only be used for green field. Because the models were pretty bad back then in science 35 days.

21:59 But uh it is it is kind of inevitable at least with software. Because it's so easy to be verified. And the quality uh the quality of the code generated is better than most people can actually hire for or buy. Now, on your topic about architecture and

2 Highlight · Using Pre-Commit Hooks as Guardrails for Loops

22:18 taste, that's what the word engineering means in loop engineering, folks. Like your job now is to actually encode a file your domain to prevent the agent from doing a commit, for example, pre-commit hooks. They're fantastic. As a human, I hate them because they slow down the ability [clears throat] to do commits. But agents don't care. So, you can make a pre-commit hook that echoes out essentially a prompt that tells it say that this boundary here can't depend upon this and that and that's just feedback. That's a feedback loop on it.

22:56 So, the engineering here is to prevent the loop from actually closing until it satisfies your engineering certification and your your requirements in the domain. So, it could be code formatting, it could be static language analyzers, it could be deterministic system testing, simulators. Like, let's put our engineering hats on. Like, we're kind of like locomotive engineers now. And it's our job to keep the locomotive on the rails. Because, um, to be frank, the model the model's a drunk. Right?

23:28 You can't trust them. But, like, we accept that, but we engineer away those failure domains.

23:34 We engineer away the failure way these failure domains. So, now is an inflection point. I guess Boris, when he first posted about Ralph back in November last year, I was like, "What the heck is Ralph?" Ralph is now, uh, it's essentially almost what, a year and a half old now?

23:54 I saw it in June 19th. It's a a year and a year and 2 weeks.

24:00 Yeah.

24:00 But you had been working on it for months at that point.

24:02 Yeah. So, it was kind of weird because we had all the YC startups all just like autonomously compressing time to build their start to build up their MVPs. And that's also something that's quite scary if you're a business founder as well. Like, if you've got a incumbent startup coming and they're building autonomously, um, and they're running much leaner and the quality it's very easy for them to actually achieve those outcomes. That adds to some of these hysteria as well

Managing Context Windows and Compounding Error Rates

24:32 because it's the topic of in business competition being at your doors faster.

24:38 All right, so I'll move on to our next question, which will be for Greg. Um, it seems like now is a large inflection point for Loops, like I said before, and compared to Jeff's announcement of the Ralph Loop a year ago, and even the widespread adoption we saw in late 2025, early 2026, um, the reason um that that may be like caught on was because maybe this is a new like capability stack where models can now process images better. They needed better verification. Um context windows got bigger and reasoning models improved.

25:07 Greg, with all of these advancements, why is the way we're using loops today still wrong in your opinion?

25:12 I mean, I don't think that model intelligence matters a lot anymore. I think it boils down to I agree with you the semantic verification, the actual ability to uh close the feedback loop, or however you call it, to actually verify that the outputs of the agent are

3 Highlight · Compounding errors in long loops

25:27 correct. And you can do it to an extent, I think. I don't think you can do it holistically, at least not at this point. I think you can do it to an extent to things that are deterministically verifiable. Uh you can get better typing in your system. You can get better linters. You can get um simulation testing and all of that. You can start keep adding that. And as long as you keep those cheap, I think that's fine. Um the moment you start adding even more non-determinism as your verification process, I think that becomes less and less correct. It starts contributing more like you know how if you prompt agent with one thing and there is a 5% chance it's going to have an error in it. And then you start looping that.

26:11 Then suddenly after 10 20 loops, it's going to be 50% chance it's correct or maybe less. And that's what I mean. And it just costed you so much money to do that. I'm going to be keep coming back to the economic viability of all of that. Um And but but to base it a little bit in like evidence, I'm pretty sure that majority of large AI companies are still using Sentry. But why is that? They they are using that just to catch simple bugs as well. It's not security bugs. It's not performance regressions, et cetera. Those problems still exist in

Context rot and the smart zone

26:41 the way that we are looping now. And we haven't solved those problems uh yet.

26:47 Thank you. For Dex, the Raffle pioneered the idea to feed fresh contacts into each iteration to avoid context rot. This has become even more manageable now that context windows have gotten much larger. Decks, are we out of the woods regarding context rot and context engineering?

27:06 I'm going to answer your question, but is there going to be like an open floor part cuz I have more questions for Jeff.

27:12 We were trying to do Oxford debate style for this one to keep it like more structured and then prevent like just

27:17 You don't want it to just turn into a chaotic yapfest.

27:21 Yeah.

27:21 to prevent what you and I do where we just start yapping. Yeah.

27:24 Yeah. It's already started.

27:25 I yeah, I was trying to control like both of you guys this time.

27:28 Okay, I'm going to do this answer as quickly as I possible and then I'm going to start busting Jeff's balls.

27:31 Yeah, you can say whatever you want with your time. You can just yeah, that's that's all good.

27:34 Okay. Uh so yeah, the the coolest thing

4 Highlight · Navigating the smart and dumb zones

27:37 about Ralph back in the day was like, okay, you keep clearing the context window and like is it completely efficient? Like probably not from a token perspective, but it meant you could leave a thing running overnight and it would never like if you just kept stuffing messages in, you would overflow the context window. But if you just relaunch it, say, here's my desired state of the world, go check the code and see what we have and do the one next step to get us there. Uh it was a very clean way to keep most of your work in what we call the smart zone of the context window. If you tell just do one thing and then we're going to clear and restart. Um context windows have gotten longer um and I will like give an update. I think I gave this in Miami, but that that video is still in production. Um The the dumb zone is really as as much as anything else is is a it's more like training wheels. Like if you have been talking to Claude for 70 70 hours a week and for two to three months, you probably don't need to think about the smart zone versus the dumb zone cuz you've built your intuition.

28:33 It's a guideline. If you're just getting This is why we teach people this is like it's a guideline if you're just getting started with AI, try to keep it around 100,000 tokens. For larger million context window, we probably don't revise the revise this up to like 200,000 tokens, but I've regularly tried to keep it under 60 for the hardest problems. I've regularly gone over 300k for things where I'm just like kind of like riffing with the agent and I'm just like too lazy to compact it and and and move on and do a new one. Um but this is your intuition. Like one of the telltale signs that you're in the that you're in the dumb zone is like uh there's certain cases where the model, you know, you're 200,000 tokens in and the model's like finished some work and it's trying to get the test to pass and it's like not working and it's like doing all these weird hacks and you read the thinking traces and it's like oh that's a test, but that's from something else and I don't need to fix that and that's a preexisting thing and you're like well no, it's not. And that that is the moment, that frustration where you're like okay, it's it's flailing trying to make something happen. That's the instinct that you a lot of people I think cultivate after a couple months working with these models. Uh but if you

29:34 don't have that yet, then you know, then then this is our guideline. So yeah, context windows are getting better. I think they're getting bigger. Um and so like the core like Ralph loop of like do as little as possible in every single iteration is like less of the motivation here than, you know, the more feedback you can pipe into the system, the more you can do autonomously and if you can have deterministic things making decisions and building small prompts to give to an agent and you don't have to remember to do that. You don't have to tell it, "Hey, go check the PR comments and fix them." and then wait and then someone makes another comment and you come back 3 hours later and say, "Oh, check the comments again."

30:07 If you can automate that process, that's great. And that's kind of the core of I think what is like loops stuff that works today. Um

30:16 Don't ask Jeff about that. That's all I have time for. I'm sorry. I have to start like really keeping this on schedule. Um okay, so now we're going to get into the anatomy of what makes a good loop a good loop. So okay, part of what makes a loop good is verification.

Preventing cheating with pre-commit hooks

30:31 However, it seems contradictory that people are saying our job is to stop writing prompts and start writing loops when the loops with bad prompts result in agents cheating and meeting its goal by by the test instead of working to pass them. Jeff, how do you keep the model from cheating when verifying its own work?

30:47 Um I heavily exploit pre-commit hooks, folks. Um and I engineer in that back pressure uh by analyzing the work that is done. Um the other thing I do is um as Dex mentioned that with Ralph, it was one of the things was everyone was trying to do compaction. Think about compaction is kind of like a lossy function, like uploading a a video to YouTube and then downloading and uploading it 100 times. Like you're losing fidelity there. And it's already a non-deterministic system, probabilistic by thing. So, the the theory behind how Ralph came to be is like, "Okay, there is a dumb zone.

31:25 And what I want to do is deterministically allocate everything it needs. Cuz if it's if it's not allocated, then it's essentially the the search space of what it can do is not constrained. But also leaving a bit of a headroom. Leaving a bit of headroom. So, I also I get meat sweats when I go above 100k, even with these million context windows. And this is really important to think about. Um a lot of people they think that you want to use LLMs um at a company.

31:53 And it's like, "I've got this data." It's like, "Sweet. Okay, you're going to have to use a loop to batch this data." What you need to think about the context windows is uh essentially remember the 720k floppy disk. You know what I mean? You've only got about a eighth of that floppy disk of usable memory that you can actually use for an LLM. So, you actually have to batch it. You can only allocate roughly around about Star Wars if you get Star Wars Episode 1 movie script and you tokenize it, you can actually just hold two of those movie scripts in memory before the context window is called. That's around about 150 kilobytes of data on a text-based movie script. So, be even very careful about this.

32:35 Something I've done it a long time and it's very silly, is uh I run a model bare um without any skills or any markdown. Actually, I get rid of all my skills and all my markdown and everything when the new model is released because the the the models actually have tastes and preferences. For example, GPT-55 when it first came out, if you screamed at it in uppercase, it became weak and timid. But if you use Anthropic, it wants you to yell at Go read the model cards, folks. Like for the integrators, like there is unique tastes for it. So, keeping it on keeping it on the rails is actually it's engineering.

33:17 It's really engineering.

33:21 Thank you.

Balancing hype and convergence engineering

33:23 Around 10 days ago, Jeff coined the term convergence engineering. He said it's where your loop slops together. Your loop uh it's where your loop slop uh comes together as a discrete like system under task until it converges. Jacks, what is wrong with how we are using loops today? How do we ensure looping slop together doesn't just produce more slop?

33:45 Uh we got to read the code. I will actually highlight uh an experiment that Jeff did earlier this year. Uh I believe what was it called? Loom? Where we had Ralph loops trying to build a software platform for the future. And I think you built you built AWS and you built GitHub and then you realized, "Okay, how do we how do we get the model feedback on things that it's not good at yet?" Like UI testing and things like this. Well, okay, the way you create a loop for is this UI good? Is you give the model something like PostHog, where it's like, "Okay, we can deploy multiple different experiments. We can see which ones the users use." And then rather than looking at screenshots and PNGs, the model can look at data and see, "Okay, uh this one is performing better.

34:27 That must be the right color for the button." And so now you've even removed like the human visual taste from the equation. And all of this sounded really cool and in the point of like how do we ensure looping doesn't bring slop together? I I don't think you can. And this is like a perfect example of the hype outrunning the discipline in the sense of Jeff, what's what's going on with Loom now?

34:49 It's still there. It's on GitHub.

34:51 Um but are you still working on it?

34:54 Uh it's been 6 months because I've been looking into engineering ways of verification.

35:00 Right. What was the thing you said to me? You said Loom's not going to work until we get better programming languages or we get better much better models. And that is a textbook for me of the hype is outrunning the discipline. We're really excited about all this stuff and by the way like everyone should do what Jeff did. Like Loom is awesome. Like go experiment, try to push the frontier cuz that's how you learn where it is and what's possible. Otherwise you just keep using your old skills with every new model and you assume it has the same limitations. But it's also a a a key point of like I don't know what I'm trying to say.

35:33 Like that this it doesn't work yet. That thing doesn't work yet. It will work someday and like there's inevitability but again it's what works today versus what is hype. So I don't know if that fully answers your question but the answer is like the way to not loop slop together and make more slop is to like read the thing that's coming out the other end and make sure it's not slop.

35:54 Yeah, that makes sense.

35:57 No, it's it's it's The labs haven't cracked it. So what makes you think you're going to crack it?

36:03 Yes.

36:04 Right? And this is right now it's we're all trying to figure out how to make this all work.

36:11 For sure. Skeptics say that loops fail quietly. They either spiral forever on your dime

Economic viability of loops in production

36:18 or the agent declares victory early on on a half-finished job. Exacts are already starting to question token spend. Greg, when does a loop pay for itself And how often is this actually the case?

36:29 I don't think they fail quietly for us.

36:33 I think they fail very very um loudly, especially when you're looking at your bills. Um, but there are cases like, as I said, I do loops, or I do engineering, I would say more so, and there are cases where I think doing loops is very valuable, or doing um or making a explicit decision that

5 Highlight · Economic viability of automated security scanning

36:56 you're going to pay pay for the cost is very valuable. So, the concrete example here is we do security scanning after on our PRs in local, and after our PRs even land, because they will always find some things that are real, um, that we have overlooked, and they sort of beat humans on the on the code review. And it's expensive. It costs us, I think, like five bucks a PR, or something like that, to run all the checks that we want. But, that's where we made an explicit decision that it's worth it. Um, there are also cases where, like, if you look at the very, very well-specified systems, such as, um, all the experiments with, uh, Next.js rewrite, or a bun rewrite in rust, or, uh, running a browser, cases where you have years and years of test suites and specifications built in around those problems, where you can really, really where well verify, um, the outcomes, then we then looping and getting to those, um, those results seems to work.

37:55 Bun in rust seems to work pretty well, from all I can tell. Um, so, there are definitely cases, and then there are cases of, um, usage that I think that I do pretty often, where you can imagine, for instance, building prototypes. I do prototypes of products that we should be doing uh doing at Sentry pretty often. Those are going to be throwaways, so I'm going to just slash gold on them, and forget about them. And if we like them, then I'm going I'm going reading the code, and I'm going to be mortified, uh, and we're going to go to the square one and start specking out what we actually meant, and and sort of go towards that solution, but it's going to be much much more involving of of human in a loop. Um so so broadly, I think they have place, uh but as Dex said, they the hype is what I have a problem with. The hype is outrunning the discipline, as he said.

38:43 And I also agree very strongly with the point that you should just try things. You should just experiment yourself, try

38:50 to see what actually works for you, where where the the cookie crumbles, um and you know, spend less time on Twitter, I think, is healthy nowadays.

39:01 Yeah, good points. A good cost-conscious loop has to track state to know what it's already tried. That memory lives somewhere on disk and get increasingly in a shared memory store that many agents read and write, especially as you go single-player to multiplayer with agents. You end up with this access control problem that can't tell which agent wrote which memory and who can read it. That might be fine for a POC, but it's definitely not for production.

39:27 The tension lies in this. Shared memory is what loops use to learn from each other and converge faster, but scoping it per agent to solve the access control problem isolates them and then kills that shared learning. Ian, how do we solve the shared memory store access control problem so loops can converge faster?

39:47 Great question. Wow. Wow.

39:51 Only if someone was working on a product that could help.

39:53 Only if someone was thinking about it. I mean, actually, I've This is This is an unsolved problem, first and foremost. Like, let's be really honest that our our access control systems weren't designed for this this world, where machines were acting and reasoning on on behalf of uh on behalf of us. But broadly speaking, I think like some of the beginnings of the substrate are starting to emerge, and if we ignore Cloud IAM and all the other stuff for a hot second, like markdown is pretty great. And so really, if we were to say a memory is markdown for the purpose of the conversation, the really question is like what things and how do I share these markdown files and use that as a memory and then how do I attribute sort of like access control around those things. And if we were to use that model, I actually think we have like the basis for for most of it today. It's just unwieldy to think about. Um a good example would be I did this tweet recently. Um I was playing with Notion and we use Notion a lot at GitHub. Uh but I really just wanted all my Notion things to be able to be as markdown files. And because it would just made it easier for the agent to work with it instead of going through

40:53 MCP and I was a bit of a CLI maxi and was looking through that. So I think what we're missing really is uh and MZ it you know Dex and I debated this last time about MCP but like we're really the challenge really is how do I present a world to an agent so that I can understand it and then how do I attribute what can access at any one point in time and how to make that wieldy for anyone to do it. And I don't I don't think we've actually cracked it but certainly there's some beginnings of patterns that make a lot of sense.

41:20 All right, now that we discussed the

Human oversight in architectural decisions

41:22 anatomy of the loop, we're going to debate the future of the loop. Like are we essentially well positioned now for software factories? Has loop engineering gotten so good that like we're ready for the full software factory? And we'll start with if loops are only good for verifiable tasks, that means fully autonomous software factories must be able to verify everything they do. Greg, is this realistic? What other parts of good engineering work such as deciding what to build, whether the abstraction is right, and what tradeoffs are acceptable?

41:50 Uh is this realistic? I think if if the compute is free, that would be pretty pretty good beginning but um

41:58 uh I think I think we're getting to the point where um or let me put it this way. The decisions that you are making as a human in in the loop are the decisions of like design, architecture, the the the important ones that um that I would say I wouldn't trust the the agent to do for me. Uh and I don't

6 Highlight · Agents naturally introduce architectural complexity

42:20 see the future where that becomes reality yet. I think the reason for that is when you're when you're looking at large organizations and I think any engineer who has had like years of experience will tell you it's not always about what you should be what you should build, but also about what you shouldn't build. What are the actual right trade-offs, where the complexity is that you want to um that you want to invest in versus where you should be um investing in maximal simplicity.

42:46 Um in my experience, agents love complexity. They will keep adding to the stack um unbounded. And so, I think we are we are shifting the the the post. We are getting to the point where as we are adding more validation, more uh semantic verification, they are able to do much much more. Um and I'm not uh not neglecting that. But I do want to be in the loop for the actual architectural decisions, and I do not see them uh taking that over any anytime

43:16 soon.

43:19 Thank you. Shubha Vice Head of Engineering told everyone at Crossover's Compile Conference that your job is just to write loops. Steinberger tweeted, "Here's your monthly reminder that you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." The view count on that tweet was 8 million people. Jeff, you've said in your original blog post, "Loops need senior expertise and it sometimes it tops out at 90% of the way there."

43:47 So, is just write the loops advice, is that safe to give to 3,000 people or 8 million people? Or only the people that have really learned to tune a loop in the first place?

43:59 Yeah, this is an interesting thing. Um it's really hard to tailor and teach like like what you should do and what you should not do to a broad that broad of an audience. So, all I can do is write to when I was originally writing, I was writing to my peers. I had people I looked up to. For some reason, I had it clicked in my head that everything has changed. The game has changed, but I was really concerned that some of the some of the best people in functional programming and like and like real peers. I remember like um Fly IO, Thomas Petek.

44:33 He he he that made him click in his head. He's got a blog post saying as such. So, I started writing for my peers and the people I looked up on. To disseminate down to like the the entire world, it's hard. Um but in the same the same sense, you go to YouTube. You got people who are creating things who've never created things before. They go to sleep and they wake up and they got a brand new Discord bot. It's magic.

45:01 I want people to remember that it is actually a pattern for allocating basically it was a pattern as an orchestration pattern. It was condensed down to a while true loop using cat cuz cat is the simplest teaching primitive. To that If you want to teach something, you got to make it really simple. Make it really simple. So, cat prompt I you engineer the prompt what it's going to be. You use the file system as state.

45:28 You recycle the context window. Run the loop. It's a little bit maybe, but it got it was just bang for buck. It just works. But, the entire attempt there is that there should be some sort of like PID controller on top or some sort of factory or some sort of some sort of some sort of determinator as such saying whether a loop should continue or not. Um it's really hard.

45:54 Um is it safe or unsafe? Safe is an interesting word. Um no one should be using coding tools on your local laptop. And this is not because of AI, but this is because of NPM supply chain attacks. This has been true. I tried to try cracking this problem for 7 years. Um and like it's now on the attention again that we that like AI could be unsafe running unsafe commands, but like your software development practices day-to-day at workplace are already unsafe folks.

46:26 So um fix that. Then these techniques start to become safe.

46:34 Thank you. Dax, you tweeted that most engineers are seeing a two to three x speed up from coding agents and that's realistic. And if you try for that 100 x speed up, you're going to get lost in the meta meta problem of optimization. And you may never get to that life-changing 10 x speed up that is possible by staying pragmatic. If only two to three x is what's possible, how do we ever get to a fully autonomous software factory?

46:58 Yeah, and I think this is it's it's a good question. I think there's like there's there's there's a mix here, too, right? We're talking about what works versus what is hype. And I think it is definitely worth I want to highlight like you should try again. Like you should try to push the frontier and do the things that might not work today, but you should not assume that all your work is changed just because you saw something work on Twitter, basically. Is like don't throw away all the things we've learned. Don't go out of your way to cast aside this, you know, decades-long career of software engineering that we as a we as a community have built up and put together. Um and it it gets back to like what I see as the biggest anti-pattern for how people set about designing and creating their software factory, which is they say, "Okay, cool. I'm going to go away for 3 months and I've read a bunch of blog posts and I'm going to go make my software factory." And it's going to be the software factory. It's going to be the future of how we ship everything.

47:59 Uh and then you come back 3 months later and you never touched the problem. You never put it in anybody's hands. It's just like any product. You're building a software factory, you're building a product for your teammates, you're building a product for 5, 10, 500, 5,000 engineers. Uh and the the right approach is to start small and iterate and figure out what works. Try things, they might not work, they might work, but the way we learn AI and how to use it effectively is through building up intuition, which is why you should try a bunch of stuff that probably won't work, but you should acknowledge that. You should not try to push like through that through that frontier. Um and so my advice is kind of like instead of trying to automate everything end-to-end, build these small incremental loops throughout your system and you will wake up one day and you will be moving two to three times faster while still being able to read the code, while still owning the architecture. And so you don't have to throw away everything we've learned and everything we know and all your intuitions just to get to this place.

48:54 And so I would I would caution people like go figure out how to make move 2x faster or 3x faster because you're going to like blow everything up by trying to go 100x faster. Um and uh yeah, I mean, if can you imagine if every software engineer in the world was two to three x faster and had a like near human 99% level human level of quality, that would change every single enterprise in the world, every single startup in the world. It would change the entire math. And we're trying to like go to Just don't don't go too far. Shoot Shoot for what you can do. Build things up iteratively. You'll learn a lot and you'll be ready for when the next models

Attribution and liability for agent code

49:31 come out and you can go 5x, 10x faster.

49:36 Thank you. When it comes to verification, it's not just about verifying the work, it's also about verifying who did the work. Employers have made it clear that humans are responsible for the code they ship whether an agent wrote it or not. Yet only one person or agent can sign a commit today. Ian, are we ready for software factories to be writing and reviewing all code if we can't determine who did the work and who the work was done on behalf of?

50:02 I think Yeah, actually I was I was playing around with this problem like two weeks ago. So, um first and foremost, we have a problem. Git only allows one signer on a commit, so we got to fix that. Um two is you know, there's some things around SOC 2 and compliance, right? But, I think like more importantly, the way to think about agents is is kind of how we think about service ownership in large organization. It's like at some point, a human has to be attributable for an agent's actions.

50:28 Like, there's no world where it's going to be like an agent is its own entity and its own attributable thing and somehow has liability. Like, the only people who have liability are people that can have consequences, right? And that always has to be grounded in it being a human. And so, if a human designs a loop and that loop presents bad software, guess who's attributable for that liability? It's going to be the human, right? Like, the society doesn't function if we don't have liability. Uh it just doesn't work. There has be consequences to damage, there's be consequences to bad decisions, and what those consequences are are obviously a gradient based on the on on the damage that's done. And if we don't have that, nothing works. And so, I'd actually broadly say there is no world where a human is never responsible.

51:06 There's always a world where a human has a level of responsibility. And the question is whether it's a human or corporation, which is a group of humans, um the question is is how does that change the way that our systems work today? So, today with Git, I can sign a commit that says I did it and it attributes to my public key, and that's cool for last generation's way that we thought about software. Now, in the future, when I tell an agent to go do something on my behalf, it generates a loop, it generates a bunch of code, it goes to production. I have to be attributable to the initial fact that I signed that thing, the that intention to do that. And we we simply do not have the substrate required um to do it, although I do think that's going to change pretty quickly. Um this isn't like something that's like crazy difficult for us to break, but we have to rethink the way that we think about attribution across the supply chain and the SDLC. And this is not a new problem, right? Like supply chain security has always been an issue.

51:57 It's actually the biggest challenge one of the biggest challenges we have with agents, and we've always wanted to have more deterministic uh pathway and sign signature chains across the supply chain, and this is just how do you do that for first-party code versus uh third-party code?

52:12 Well, Ian, um, to to just to shoot straight, um, our profession is a bit of a clown show. We actually don't have liability at a personal level. Like we call ourselves engineers, we're not really engineers.

Final panel insights on types and supply chains

52:24 Some of this stuff is going to get really complicated, folks, and maybe we need to revisit these topics.

52:31 And now that we've had our main debate section, to wrap it up, each member on each team will get 2 minutes to uh wrap it up and describe their final thoughts before we decide who the winner was. Uh, Greg, what would you like to

52:45 Yeah, absolutely. Um, it's it's interesting seeing how many how many points we actually agree with each other, uh, but that's how how this goes. I I do think that the biggest point that I'm trying to I have been trying to drive here is um where we are whether we are where we are trying to be and then what you're actually hearing from the ever-present hype uh hype loop. And that's where I want to sort of double down and tone this down. I basically want to say try things. Think for yourself. Don't lean into the bubble. Uh, don't lean into the hype because you will find what works for you and you will find where the system breaks the best by doing it yourself. That's always has been how um how humans learn the best is by practice, not by, you know, watching YouTube. Um, and ultimately, that's that's what I'm trying to do, and I'm slightly skeptical uh when it comes to, you know, allowing the full loop to run because I see the qualitative results not being up to snuff for my

53:46 my requirements. But, I am optimistic in general. I am optimistic because I've seen how much more I can do and what are the types of problems I can address today that I couldn't address even a year ago, let alone let alone earlier than that. And and it's going to get better, but I'm also not worried about my software engineering career. I don't think we're going away. I think we are still going to be important piece of this whole software factory or whatever the next um sort of hype bubble is going to become.

54:18 Ian, would you like to go next?

54:19 I would love to. So, the train's left the station. This stuff works, um and there is a real productivity increment, right? Like, more stuff is getting done. It doesn't mean all the stuff getting done is good, but more stuff is getting done and a good percentage of that stuff is actually good, right? I think we can agree that. And second is now we have competitive dynamics, where it's no longer possible for a company to sit back like, "Hey, we're going to sit this loop stuff out, right? I'm going to sit this coding agent stuff out." Like, that that's over. We're we're past that.

54:48 Like, that train leaves the station. As soon as that train leaves the station, everybody in the world starts saying, "Holy I got to keep my stuff together. We're going to stay up with the Joneses." And they have to because we live in a very competitive capitalist society, and that's also, you know, ducking politics, but it's it is what we are. And at the end of the day, the question is is not really um will it happen, it's when it happens and to what degree over time. And I don't think there's a choice but to actually stay up-to-date, and we're all just kind of holding on to a rocket ship ride, where we don't actually know what that We don't really know the trajectory other than it feels real fast with crazy acceleration, and sometimes like we're way ahead of each other, and sometimes we're way behind, but I don't think you have a choice not to be one, figuring out what loops are, two, figuring out where you can apply them in your code base. There's going to be places where, "Hey, this is highly verifiable. This is a problem that like computers can solve.

55:38 It doesn't make a lot of sense. Cough connectors is a good place of like, you know, last 10 years. The amount of companies that have made money because they're basically just connector farms, there's no longer a lot of value there, right? If you can automate the connection connector creation. So, there's places in software that you can apply loops today and get real value and there's places where you probably shouldn't and you should decide where that is and it's probably the core value of what you're creating, right? And so, that's how I think about it. Broadly speaking, trends left the station. The productivity curve is what drives society.

56:09 The ROI, so how much more productive we are as a society is what drives GDP. What drives GDP is ultimately where the dollars go um and how capital gets allocated.

56:18 Okay, get over to you next.

56:20 I eagerly await the world where the life soft software factory is feasible. The I I would love a world where we don't have to read the code, where where we can just do everything. Um if you watch my talk on Tuesday, I I think this is actually a problem we can only solve at the model level right now. I don't think the harness can do it uh unfortunately cuz I love building harnesses and doing context engineering. Um so, my advice is uh pay attention to Jeff.

56:48 Uh let me know when Loom is actually working and until then uh use loops, but not like that.

56:57 Love it. So, so much to say, folks. Um factories represents where we're heading to the future. Like it it's it's essentially like a perpetual energy motion machine, right? It's it's like it's the pipe dream. Companies are only just getting founded today and getting receiving their rounds today. Don't think you can just take this and implement it in your company cuz it's just generally not solved in market.

57:27 But, I will say to add to the meta monologue, if you try to run in loops or try to build a factory to using Python, it's going to be a clown show.

57:39 If you do it in Ruby, it's going to be a clown show. Static types are a form of verification, folks. I encourage you to to come up with a couple of cutters and a couple of experiments. Try running some loops. Build an application in Ruby and then try to modify it again with these loops and you'll see the maintainability mess. And then try doing it with Haskell. I don't care if you don't know Haskell. It doesn't matter if you don't understand Haskell. The LLM understands Haskell and you can actually prompt the the agent to explain this as if you to to your son or daughter.

58:15 So, I'm not sure even code needs to be readable these days, but this is frontier frontier thinking. It needs to be explainable. So, I'm playing around with different

7 Highlight · Minimizing supply chain risk via code generation

58:24 domains of verification here, but uh one thing's for sure, um type systems are in. Very much in. Rust is very good because how the ecosystem models some types. And because supply chain was mentioned, I need to say it has been 10 months now. I do I minimally use any open source software. Minimally. I just generate it to my requirements. And then when a supply chain attack happens, I'm like, "Didn't affect me."

58:55 All software has security flaws, but it's about minimizing the blast radius. And the thing is, if you deal with open source project, the person's on on leave, the maintainer, what have you, you can't talk to a human. That's not AGI. You want to be venturing all your source code as much as possible so the so the agent can actually modify and in without a loop or other prompting. You need to

59:19 own your supply chain.

59:21 Awesome. Thank you so much for that. I personally have to close it at that, but we can now decide our winner. Um Show of hands in the audience if your mind was changed and you're now on Dex's side, raise your hand.

59:37 Let's go.

59:38 Okay. If your mind was changed and you're now on Ian and Jeff's side.

59:43 Yeah.

59:44 Yeah.

59:45 I kind of couldn't see with the light. What did you guys think? It was pretty good. It's pretty close, yeah.

59:52 It's impossible. The lights are so bright.

59:54 All right. Well, that was a good debate. Thank you so much for listening. We'll see you next time.