My own version of OpenClaw focused on 24x7 development

Speaker 0: We do stuff. And yeah. So we have a little project that's arguably that's, you know, high 6, low 7 effort. Right? And I'm trying to build exactly that. Speaker 0: But we're not with the client's name, not with exact spec, not with exact same requirements. The idea here is that I want to prove to my client that my harness, right, could potentially cut the cost of whatever if you produce the perfect selling requirements, nonfunctional and functional being fed in. Instead of building it the traditional way, right, of hiring people to build it, you build a harness and you would destroy the output. Instead of fixing the output, you would again fix the input or fix the harness until you're 90%, 95% there. And then you have AI driven engineers to tweak it with AI. Speaker 0: Right? So that's my goal and that's what I'm trying to show my clients and it's about that that dollar range. Speaker 1: How do you how do you keep your agent aligned over that? Speaker 0: Great question. Do you like verifiers or something that Speaker 1: you took place or Speaker 0: Yeah. So being transparent, that's why my harness is better than my 24 hour agent. My harness is better doing that. Right now, I'm not doing a good job of doing that whatsoever. But it is open source. Speaker 0: So it is, Rob Mike me asked me to open source this. I think it's open source and I, you know, I will continue to build on it. If anyone wants to try it out, feel free. The code is out there and hopefully, yeah, someone checks it out. Sean. Speaker 0: Good question. So what I do is I have a retrospective thing. I literally do a retrospective with my agent. I tell her what's wrong and probably what was wrong. So we kinda both kinda came up with what's could do what was done correctly and what could be done better. Speaker 0: And that gets fed into either improving the input. Right? Remember, they're they're kinda like the prompt packet. Right? Mike, the many, many, many tokens of large sized things. Speaker 0: You you gotta add more details to it, right, so you're more consistent, and also fixing the both executive agents and the work agent skills. I don't wanna be very skillful with this right now. Like, Mike my first 1, no skills. Now I'm all skills. I'm only on skills. Speaker 0: So if the skill is not right, like, we need to fix the skills. And so that next time we run, AI has a better idea what it's doing, so I have a better product. Alright. Thank you so much. Speaker 1: And I was adjusting real time. We actually only have 4 talks, so, Sean, you're next. Okay. So I was giving a little more time for questions and everything. So, alright. Speaker 1: So my talk is on reverse Ralph loops. How many people know what a Ralph Wiggum loop is? Okay. Good. There's some people in the room that know what this is. Speaker 1: I put this up here because this essentially demonstrates the foundations for what that is. You define what your requirements are. You create an implementation plan, and then you implement and just let it let it put the element in a loop. So I've been doing this, I don't know, since last year sometime. Jeff Huntley, that's what this website is. Speaker 1: He put this out in July, but it's very good for greenfield. It still works for brownfield, but you have to know how to wield it. But this really started to work in December, and that's why you Mike have seen a lot on x and Twitter in January because it just started working with all the November models. So just putting it in a loop. Well, what I've discovered is a lot of open source is going to start going closed source. Speaker 1: They are going they already started, like, hiding their tests for open source because you can reproduce any software that you want right now, and all you have to do is put an LLM in a loop. So I