WEBVTT

00:00:00.000 --> 00:00:22.480
Speaker 0: I came across this post by Peter Steinberger, the creator of Open Claw course yesterday on X where he wrote So loop engineering.

00:00:22.540 --> 00:00:23.680
Speaker 0: it is now, right?

00:00:24.080 --> 00:00:27.600
Speaker 0: I don't think its an official term yet but we'll see if will be.

00:00:27.920 --> 00:01:05.160
Speaker 0: And of course were coming from a past where we had prompt-engineering then parts decided that this should be rephrased or relabeled as context engineering which was always stupid because the same thing in end and because it ALWAYS WAS about ensuring that model has right context matters, did matter still matters will matter because if you want to have better chances of getting good results out You have a better chance than no guarantee.

00:01:05.360 --> 00:01:08.100
Speaker 0: even with the right context mistakes are possible.

00:01:08.160 --> 00:01:10.320
Speaker 0: We're just not getting what you were looking for.

00:01:10.480 --> 00:01:23.460
Speaker 0: that's all possible because it still A non deterministic system and on deterministic tool, but if you want to have a shot at getting good results And you definitely can get good results then providing The right context is important.

00:01:24.040 --> 00:01:42.380
Speaker 0: now around Change from twenty-twenty five to twenty-two six and of course throughout this year We then saw the rise of agent decoding since tools like Claude code and Codex Combined with the models that are used inside them which have been heavily fine tuned an optimized for Instruction following in coding tasks.

00:01:42.960 --> 00:01:45.100
Speaker 0: those tools with the model showed us.

00:01:45.160 --> 00:01:58.620
Speaker 0: That yeah, you can really use these AI models LLMs For coding tasks and get stuff done with them as assistants, at least that is still my take and experience.

00:01:58.680 --> 00:02:06.160
Speaker 0: And I've been using these models a lot in these tools playing around with them pretty much every day... ...using them everyday!

00:02:06.620 --> 00:02:09.860
Speaker 0: Not just playing around for serious projects.

00:02:10.139 --> 00:02:17.540
Speaker 0: Of course this is why i built courses about cloud code and codex where i dive deeper to share my learnings on how to use those tools.

00:02:18.720 --> 00:02:33.120
Speaker 0: These are useful assistance but they just aren't those replacements of developers yet, and as I've shared in many other episodes probably also not in the near future.

00:02:33.800 --> 00:02:55.880
Speaker 0: Nonetheless, of course, Anthropic and OpenAI added extra commands to these tools like the slash goal command in Codex or the slash loop command in Cloud code where you can specify a specific goal a maybe more complex task with that command added in front of it.

00:02:56.460 --> 00:03:06.020
Speaker 0: And the tool codex, Cloud Code with the model will keep on going and we'll keep on re-prompting itself until that task is completed!

00:03:06.460 --> 00:03:10.060
Speaker 0: It's kind only the Rolf loop again remember?

00:03:10.160 --> 00:03:10.900
Speaker 0: The Rolf Loop.

00:03:10.980 --> 00:03:24.020
Speaker 0: at the beginning of twenty twenty six We had that hype around the Rolfloop where some people just sold you That You Just Need A Detailed Step By Step List of tasks that need to be completed.

00:03:24.160 --> 00:03:34.580
Speaker 0: To achieve a certain goal, build a certain feature and then you could use an extension... ...to keep cloud code and code extend at some point going in and work its way through the list.

00:03:34.660 --> 00:03:46.160
Speaker 0: And even though we had the RALF loop back in January already.. ..and some people sold it as the solution for building software autonomously Where is all that software?

00:03:46.280 --> 00:03:50.280
Speaker 0: where Is All That Software, that error-free amazing software.

00:03:50.360 --> 00:03:52.340
Speaker 0: Why is Claude code still flickering?

00:03:54.120 --> 00:03:55.820
Speaker 0: Yeah Anyway.

00:03:56.620 --> 00:03:58.600
Speaker 0: So we had the raw loop back then.

00:03:58.900 --> 00:04:02.980
Speaker 0: now it's back here officially integrated into Claude Code and Codex.

00:04:03.100 --> 00:04:09.980
Speaker 0: And Now We're Talking About Loop Engineering or Designing Your Loops That Prompt your Agents.

00:04:10.120 --> 00:04:16.399
Speaker 0: of course there is something thats easy to say for someone Who works for open AI.

00:04:17.000 --> 00:04:19.920
Speaker 0: in the end because of unlimited tokens?

00:04:20.240 --> 00:04:25.700
Speaker 0: Because it turns out this these loops, these commands they can burn through a lot of tokens.

00:04:27.260 --> 00:04:52.800
Speaker 0: The problem just is you have to same probabilistic nature off the entire system and I think one thing that's often overlooked is, indeed my experience has been these AI models and the tools combined are pretty good at keeping on going until a certain goal is achieved.

00:04:58.500 --> 00:05:03.180
Speaker 0: months ago now, is I had a couple of pdf documents which i needed to combine into one.

00:05:03.280 --> 00:05:10.100
Speaker 0: Which combined must not be bigger than five megabytes but each individual document was already like six megabytes because they contain scans.

00:05:10.580 --> 00:05:19.600
Speaker 0: so just threw my coding agent codecs at the task and it kept on going until it really achieved that!

00:05:19.960 --> 00:05:21.100
Speaker 0: And obviously this might.

00:05:22.800 --> 00:05:31.100
Speaker 0: The point just is, indeed these models if they can verify an outcome are quite decent at achieving a goal.

00:05:33.720 --> 00:05:37.420
Speaker 0: They keep on going and try different ways of getting there!

00:05:37.920 --> 00:05:45.700
Speaker 0: That problem isn't necessarily how good software is being built – it's one thing to get something done….

00:05:49.700 --> 00:05:51.900
Speaker 0: that may be enough for certain use cases.

00:05:52.320 --> 00:05:58.000
Speaker 0: if we're talking about software that should be distributed, evolved and maintained.

00:05:58.800 --> 00:06:08.120
Speaker 0: It's not a good strategy to just find A way of getting there because THAT ONE WAY MAY GET ONE THING DONE at this point in time!

00:06:10.000 --> 00:06:15.020
Speaker 0: IT may break for slightly different input it may contain a lot bugs or security issues.

00:06:15.760 --> 00:06:20.340
Speaker 0: it may fail for so many reasons, for so much other situations.

00:06:20.420 --> 00:06:27.560
Speaker 0: It may have poor performance and all that again might not matter if you're just trying to get one thing done right now.

00:06:28.400 --> 00:06:35.560
Speaker 0: but that is again NOT what software in general... ...if we are talking about software as a product at least....is about!

00:06:36.120 --> 00:06:48.780
Speaker 0: So there ARE reasons why we learned as developers that certain patterns and practices and approaches make sense because they're easier to adapt, easier to understand.

00:06:48.940 --> 00:06:52.860
Speaker 0: Easier to adjust simply cleaner.

00:06:53.120 --> 00:07:01.980
Speaker 0: not just for the cleanliness sake but For the extensibility maintainability performance security And understanding ability.

00:07:02.140 --> 00:07:17.840
Speaker 0: say and even if you don't care about Understanding The code anymore Because You'd Say That AI just needs to understand it not a human which is all the really, really bad take Because obviously AI models have limited context windows and all that.

00:07:17.900 --> 00:07:21.460
Speaker 0: but even then if that's your take on the Understandability.

00:07:21.900 --> 00:07:24.320
Speaker 0: The other parts still matter.

00:07:24.740 --> 00:07:30.180
Speaker 0: And Yeah I don't think there is more to say about dad II.

00:07:30.300 --> 00:07:46.340
Speaker 0: Really hate the current point in time where we have all these annoying stupid terms coming up all the time and then we got people trying to sell you products, courses.

00:07:47.200 --> 00:07:52.160
Speaker 0: I just don't sell your course on loop engineering or anything like that.

00:07:52.840 --> 00:08:10.980
Speaker 0: but yeah here coding agents for what they are helpful assistants, but right now we're still stuck here and I'm excited to see What will be next after loop engineering?