Transcript
00:00:00I came across this post by Peter Steinberger,
00:00:02the creator of OpenClaw, of course, yesterday on X,
00:00:06where he wrote, "Here's your monthly reminder
00:00:09that you shouldn't be prompting coding agents anymore.
00:00:12You should be designing loops that prompt your agents."
00:00:17And oh boy, I have some thoughts here.
00:00:21So loop engineering it is now, right?
00:00:23I don't think it's an official term yet,
00:00:26but we'll see if it will be.
00:00:27And of course, we're coming from a past
00:00:29where we had prompt engineering.
00:00:31Then parts of the industry decided
00:00:33that this should be rephrased or relabeled
00:00:37as context engineering, which was always stupid
00:00:39because it's the same thing in the end,
00:00:41because it always was about ensuring
00:00:43that the model has the right context.
00:00:45That was the entire idea behind prompt engineering too,
00:00:48because yeah, obviously the right context matters,
00:00:51did matter, still matters, will matter,
00:00:54because if you wanna have better chances
00:00:57of getting good results out of LLMs,
00:00:59you need to give them the right context.
00:01:02You have a better chance then, no guarantee.
00:01:05Even with the right context, mistakes are possible.
00:01:07We're just not getting what you were looking for.
00:01:10That's all possible because it's still
00:01:12a non-deterministic system, a non-deterministic tool.
00:01:15But if you wanna have a shot at getting good results,
00:01:18and you definitely can get good results,
00:01:20then providing the right context is important.
00:01:23Now, around the change from 2025 to 2026 and of course,
00:01:28throughout this year, we then saw the rise of agent decoding
00:01:32since tools like CloudCode and Codex combined with the models
00:01:36that are used inside them, which have been heavily fine-tuned
00:01:39and optimized for instruction following and coding tasks,
00:01:42those tools with the models showed us that, yeah,
00:01:45you can really use these AI models, LLMs for coding tasks
00:01:51and get stuff done with them as assistance.
00:01:55At least that is still my take and my experience.
00:01:58And I've been using these models a lot and these tools,
00:02:02playing around with them pretty much every day,
00:02:05using them every day and not just playing around with them,
00:02:07also using them for serious projects.
00:02:10And of course, that is why I built courses
00:02:12about CloudCode and Codex, where I dive a bit deeper
00:02:15and share my learnings and how to use these tools.
00:02:17And these tools are useful assistance,
00:02:21but they just aren't those replacements
00:02:25of developers yet.
00:02:28And as I've shared in many other episodes,
00:02:31probably also not in the near future.
00:02:33Nonetheless, of course, Anthropic and OpenAI,
00:02:36they added extra commands to these tools
00:02:41like the /goal command in Codex
00:02:43or the /loop command in CloudCode,
00:02:46where the idea is that you can specify a specific goal,
00:02:51a maybe more complex task,
00:02:53with that command added in front of it.
00:02:56And the tool, Codex, CloudCode with the model,
00:03:00will keep on going and will keep on re-prompting itself
00:03:03until that task is completed.
00:03:06And it's kind of only the RALF loop again.
00:03:09Remember the RALF loop at the beginning of 2026,
00:03:13we had that hype around the RALF loop,
00:03:16where some people just sold you that you just need a detailed,
00:03:19step-by-step list of tasks that need to be completed
00:03:24to achieve a certain goal, build a certain feature,
00:03:27and then you could use an extension
00:03:28to keep CloudCode and Codex then at some point going
00:03:33and then work its way through that list.
00:03:34And even though we had the RALF loop back in January already,
00:03:38and some people sold it to you as the solution
00:03:41for building software autonomously,
00:03:44where is all that software?
00:03:46Where is all that software, that error-free, amazing software?
00:03:50Why is CloudCode still flickering?
00:03:54Yeah.
00:03:55Anyway, so we had the RALF loop back then.
00:03:58Now it's back here, officially integrated into CloudCode and Codex.
00:04:03And now we're talking about loop engineering
00:04:05or designing your loops that prompt your agents.
00:04:10And of course, that is something that's easy to say
00:04:12for someone who works for OpenAI in the end
00:04:17because of unlimited tokens,
00:04:20because it turns out this, these loops, these commands,
00:04:23they can burn through a lot of tokens.
00:04:27The problem just is you have the same probabilistic nature
00:04:33of the entire system.
00:04:34And I think one thing that's often overlooked
00:04:38is that indeed my experience has been
00:04:40that these AI models and/or these tools
00:04:43and the models combined, it's really both.
00:04:46They are indeed pretty good at just keeping on going
00:04:50until a certain goal is achieved.
00:04:52I mean, one tiny example I had a few weeks or months ago now
00:04:59is I had a couple of PDF documents
00:05:01which I needed to combine into one,
00:05:03which combined must not be bigger than five megabytes,
00:05:06but each individual document was already like six megabytes
00:05:08because they contain scans.
00:05:10So I just threw my coding agent, I think Codex at the task,
00:05:14and it kept on going, kept on writing some little programs and stuff
00:05:17and until it really achieved that.
00:05:19And obviously, that might not be a super complex task.
00:05:22The point just is, indeed, these models,
00:05:25if they can verify an outcome, they are quite decent at achieving a goal,
00:05:30at achieving a certain task.
00:05:32They just keep on going and try different ways of getting there.
00:05:36The problem just is that is not necessarily how good software is being built.
00:05:41It's one thing to just get something done, to just find a way of doing something.
00:05:48That may be enough for certain use cases.
00:05:51If we're talking about software, software that should be distributed,
00:05:54that should be evolved and maintained,
00:05:57it's not a good strategy to just find a way of getting there
00:06:02because that one way may get one thing done at this point in time.
00:06:08It may break in the future.
00:06:09It may break for a slightly different input.
00:06:11It may contain a lot of bugs or security issues.
00:06:15It may fail for so many reasons, for so many other situations.
00:06:20It may have poor performance.
00:06:22And all that, again, may not matter
00:06:24if you're just trying to get one thing done right now.
00:06:28But that is, again, not what software, in general,
00:06:31if we're talking about software as a product, at least, is about.
00:06:35So there are reasons why we learned as developers
00:06:41that certain patterns and practices and approaches make sense
00:06:45because they're easier to adapt, easier to understand, easier to adjust.
00:06:51Simply cleaner, not just for the cleanliness sake,
00:06:55but for the extensibility, maintainability, performance, security,
00:07:00and understandability sake.
00:07:02And even if you don't care about understanding the code anymore,
00:07:06because you'd say that the AI just needs to understand it, not a human,
00:07:10which is all the really, really a bad take,
00:07:14because obviously AI models have limited context windows and all that.
00:07:17But even then, if that's your take on the understandability,
00:07:21the other parts still matter.
00:07:23And yeah, I don't think there is more to say about that.
00:07:29I really hate the current point in time where we have all these annoying,
00:07:37stupid terms coming up all the time.
00:07:39And then we got people trying to sell you products and courses and stuff off that.
00:07:45And I sell courses myself.
00:07:47I just don't sell and won't sell you a course on loop engineering or anything like that.
00:07:52But yeah, here we are.
00:07:54I'm sure at some point we'll be past that.
00:07:58And we can use these coding agents for what they are helpful assistance.
00:08:03But right now we're still stuck here.
00:08:05And I'm excited to see what will be next after loop engineering.
Community Posts
No posts yet. Be the first to write about this video!
Write about this video