The OpenClaw Podcast - The ClawCast - Episode 5

[00:00:02]Hello everybody and welcome to the clawcast the official open claw weekly podcast and today I am joined by my co-host uh Kevin my co-host is Patrick my guest is Kevin Patrick is with the open claw foundation uh the dev team and Kevin is from openai welcome both of you >> thanks excited to be here >> yeah glad to be So, I've got a few questions to get things rolling. Uh, we we sort of deferred this discussion a few times due to the release of 5.6 being delayed and then you had a vacation and etc etc. So, uh can you tell us what exactly you do at OpenAI? What is your role?

[00:00:50]>> Yeah, uh this has changed over time. So I joined OpenAI uh around two years ago um from uh another startup that I founded at OpenAI. I started off in in the info team. So worked on building up our own info stack and um yeah I'll I'll leave it at that. Um afterwards I led our connector products team. So this is essentially whenever you use connectors in chat GPT they're now called plugins. So, Gmail, uh, notion, uh, or whatnot. Um, and then around two months ago, I ended up teaming up with Peter and essentially, uh, we have formed a team around OpenClaw at OpenAI and, uh, currently the OpenClaw enterprise efforts, uh, which has a lot of intersections with also OpenClaw, the open source project.

[00:01:45]>> Yeah, cool. Um that's very exciting. So you do work while working for OpenAI on Codex. You you do work on OpenClaw.

[00:01:57]>> Yeah, I'm actually one of the maintainers. Uh so are a couple of the other people here at Codex. Um when we first started claw labs, which is uh the team at OpenAI on OpenClaw, uh most of the work, almost all of the work essentially was on the open source. So this was making sure that the when you're using OpenAI models that they didn't suck. Um and also uh integrating the codeex harness uh into OpenCloud when you're using Open AI models which means that you get a lot of things for free like auto review uh like being able to use your um OpenAI plugins uh by default and all sorts of other really cool things from the Codex harness. The harness piece I I think is one of the things that still isn't quite as commonly known as I would expect it to be that like openclaw itself is harness agnostic whether you want pi whether you want codeex in the future other you know harnesses that would like to integrate but like you said there's just so many advantages of that and I think it's it's a good example of just how sort of like Switzerlandy openclaw is about every single layer of the stack and what that stack is kind of emerging to be right like the harness itself is just kind of one piece of the whole thing.

[00:03:08]>> Absolutely. And I think that's also a place where like I'm constantly explaining to people when talking about OpenClaw because you know I think the great promise of OpenClaw is just it's this open platform where everything is customizable including the harness and then I got this funny look of like wait isn't openclaw a harness? What do you mean like that's >> it's no it's like I think the harness is a part of openclaw but openclaw itself I see as an end toend platform for you know building agents and so if you think of it as a car the harness is an engine but then there's everything else around it and I think openclaw is yeah it's that it's an end toend car if you will.

[00:03:48]>> Yeah. Um I guess the the reason we delayed you joining podcast was because of 5.6.

[00:03:58]Um it is a really really great model.

[00:04:01]Can you tell me more about more about it? More about what you guys have been hearing since it's been released. What has what what have been the biggest surprises about its use?

[00:04:12]>> Um yeah, so 516 uh we've been using it for a while now. It's it's really great.

[00:04:19]Um, it actually made it really difficult. I think one of the it's a nice problem if you have that is just like when you're switching from your work account back to your personal account. Um, it it feels like a major downgrade whenever you go back to like a public model.

[00:04:31]>> People get kind of like depressed here internally where it's like what is even the purpose of my work if like you know 56 could have one shoted this and I'm just sitting here you know doing its work for it. Like I hear it's kind of depressing in a way. Um [sighs] yeah, I mean I think it's like it used to be kind of like with computer hardware cycles where it's like you know should I buy like the latest computer now or should I wait in 2 years where like the hardware is going to be exponentially better and cheaper and this is a similar curve with uh models. Uh I definitely have friends at the company who are like you know I have these projects I want to make but like what's the point because I can just you know wait a few months and then the next model is just going to be able to oneshot it. Um but I think >> maybe like nihilism is a better term than depression around the sort of the feeling.

[00:05:19]>> Yeah. And I think my take on that is just like you know like at the end of the day you're living in the present. So it's like whatever like you need to pull forward whatever you want to uh make now. Um but I guess like getting back on the point of like 5.6 uh yeah 5.6 is an incredible model. I use it as my daily driver for basically everything. Um it is really good at long horizon tasks. It's basically I know like before 5.6 um I was a big fan of slash goal in codeex um which essentially you lets you set up a R loop um where the model would continue iterating on something until it accomplish the goal um with 5.6 six that basically is not necessary because the model has been trained to essentially continue working until it has accomplished a goal. Um so I think that's like you know one of the defining traits in 5.6 is that it is really persistent at finishing a task. Sometimes maybe to the detriment of uh uh other criteria, but um yeah, it's it's really good at at uh finishing complicated tasks. It's really good at planning. Um, it's very token efficient in in the sense that like it it might use more tokens because it's more comput intensive, but it usually uses fewer tokens overall on complicated projects. Um, and all of this I mean I would say like all of this is also a direct result of the post training like these were explicit goals for 5.6 six that it is much more token efficient that it is good at >> long horizon task and then also like much better at computer use and multi- aent work.

[00:07:12]>> Oh, so good at computer use. This morning I had it um actually reconciling my uh bank account in QuickBooks. It was doing a great job. So >> I I've been using 56 to help my parents with kind of like we were talking about Hannah. They have like a small business where I've been running like their Google ads for them and there's something where like I had to email them and then they pass me to another email chain and I've just been having 56 cuz Google ads does not have like CLI or like any good support. So it's like 56 is just driving the like ridiculously complicated Google ads council for me to like figure out where to submit my help tickets and and things like this. But um Kevin, you mentioned like 56 for like daily driver. I'm assuming like 56 soul with like high or x high for like coding tasks. But um when it comes to like kind of broader set of like personal agent like open claw adjacent tasks do you find uh like for you personally I guess like are you running soul are you running like era for more you know these kind of you know like day-to-day type tasks and maybe just generally too like I honestly don't think I have a great understanding of like when to split a model into like a soul terra Luna versus just like you know keep like a single one through five you know thinking slider so maybe some of that too.

[00:08:33]>> Yeah. Uh I think the model picker and intelligence levels is just this giant pile of complexity. Uh we tried to kill it with GPD5. Um and instead of killing it, it is now back more than ever. Um you have different models, different thinking levels. Um, and I don't think there's like a one good way of thinking about it. Everyone kind of has a different heristic.

[00:08:59]For me, I generally start off most things using 5.6 sole medium. Um, I think that's become more or less the consensus of medium's a good place to start off because it's good for most tasks and switching to high you do have like a bump in token usage and compute which is not always necessary. Um, and yeah, I use 5.6 medium as default for my open claw, for example. Um, I would say like for terror and Luna, um, I would say like Luna probably as for more like really go like I need to look up a bash command or like, you know, here is something that uh, uh, needs to run really fast because uh, latency wise uh, it can be better. Um that said I will admit like one of the perks probably the best perk of working at a frontier lab is you have unlimited tokens and so it does mean um for most things you know I will just oh I'll you know use 5.6 or whatever the latest one is because um infinite tokens and whatnot. Um but that said like this show like a real like latency uh issue for if I want something to be done faster I'll use a smaller model. Um, and yeah, I guess this is also a funny story of like initially when I was working on OpenClaw, um, OpenClaw wasn't actually allowed at OpenAI. Um, which meant I couldn't develop an OpenClaw using my corporate account, which meant that I actually had to use a regular pro account. And then really feeling the token burn, I ended up creating like five pro accounts to just cycle between because I kept on hitting my limits. um you know uh different habits I guess but um yeah that was just interesting.

[00:10:50]>> Well, somebody just asked they said uh I'm curious to know if you know um who uses OpenCloud OpenAI not for testing and what use cases they have uh since you guys are all using codecs and then after that um Patrick we can segue into what we were talking about earlier regarding coding.

[00:11:12]So >> yeah, I mean in terms of Open Claw and OpenAI, um this is pretty pertinent because like this is what my team currently is in charge of is how do we roll out OpenCloud to the enterprise in a way that is both useful and safe and as part of that like we're also piloting that at OpenAI. Um so you know right now like we do have claw adoption inside of the company in an official capacity. I would say clause in general there's two general categories that they fall into. It's either personal clause or team clause. So personal clause is it's my executive assistant. Um you know help me schedule uh like a calendar meeting with somebody that I'm having a conversation with or like add additional context. uh if I'm in a conversation um and somebody is asking for like a project they worked on like using my personal claw to like fill them in on the details if it's like pretty trivial to do so um versus like a team claw like I think team claws are also really interesting um in it's you can think of team class as more like actually getting into work automation and being a persistent teammate um so like examples is like oh like the security team like using security claw to do the first pass at doing security reviews or like the goth team using like a ghost claw to essentially help them run goth experiments. Um I think there's a place for both. Uh I I would say like definitely believe in this future where we're moving into this work environment where it's going to be teams of agents working with teams of people and what that interaction model is going to be like is what we're trying to uh currently model. I I I actually do think a lot of that is going to look like, you know, people commuting communicating with um something like a claw in channels like Teams and Slack just because that's currently for better or worse where the work happens right now.

[00:13:18]>> We were just talking about before the call as well the new uh product from Block that's kind of their like agent specific version of like Slack and these things that's open source. So I think like yeah everything that you said I feel like is really clear and then I feel like there's this whole other sort of layer of like what is actually the right sort of like product interface to like deploy a claw to especially when like a lot of these a lot of companies would prefer to keep their their data like locked in so that like you know you use like their agentic product where it's limited to their box and you can't you know take the data and bring it to where you want.

[00:13:57]Um yes and was there a question there or >> no no just a amusing >> yes >> there is um coding in open claw because people often say what's the difference between codeex and openclaw why would I use codeex or openclaw instead of codeex and quite frankly um currently why would I code with openclaw because it sort of sucks at it and I' i'm dropping that sucks line again. But um that brings me to the point where this month uh one of our internal goals is make it so that we can code first from you know build openclaw with openclaw and you know under the surface if we're using open AI models um which which we regularly do uh then it's still using codeex app server but this is something that uh is is coming up on our radar and we really want to do how does that impact you guys if you guys if you're using an open claw a claw in your setup is there a chance that you'll help to I guess basically orchestrate your work to codeex through through a tool like openclaw yeah uh I mean I think right now like anything can work um as you pointed out like if you're using an openi model with openclaw today you're already running the codeex app server uh which is the codeex harness like this is what the codeex app uses is this is what you get when you use the codec cli. Um there are some differences because when you're running the codeex artist inside of openclaw um it we uh a some of the tools get shimmed. So you're using codeex tools for some but then you're using openclaw tools for other things and um you might have like slight slight tweaks in behavior. Um but you're still using codeex at the end of the day. Um all I can say is like when I am like launching uh coding jobs from openclaw it's generally telling openclaw to tell codeex to execute something. Um so openclaw becomes essentially an orchestrator of work. Um and it's not necessarily just um like codecs for example like in my personal use like I regularly experiment with other models.

[00:16:18]So I also have an entropic subscription as well as codeex and so [snorts] um it's you know using open cloud to orchestrate both cloud and codecs on different jobs.

[00:16:30]>> Yeah. Yeah. I think for me personally like the when it comes to just having like a very high level conversation of like here's something I want to work on and like maybe I'm on like a walk for example I kick it off to openclaw and then it you know delegates to codeex like like openclaw has for a while worked well for me in those scenarios and then to our point about like the harness it's like this is running actual codecs like it it's it's as good at writing code as codecs because it is running Codex underneath the hood. But where I think we as a team have not been able to like dog food or as as some of my friends like to say like eat your own pizza for something a little better. Uh use code or use open cloud to build open cloud. The problem has been like when you really want to like dive in and actually kind of like pair with an agent on like the the more granular pieces of the work instead of just saying like hey go do this then like is it an IDE? Is it like a aentic development environment? whatever you want to call like effectively the codeex app like you need some ability to like pair with the agent and like the control UI and openclaw up until like 72 this upcoming release I'd say has just not been like a a good enough or like a polished enough piece of software for me to like daily drive this thing um and that's like a lot of the the control UI stuff that that'll be coming out in 72 uh 2026 72 release of open clause our attempt at like driving that forward to the point where like we can actually like you know start in like a you know any given channel for that matter just to give a prompt but then when it comes to like actually drilling down and pairing with an agent on a piece of work the the control UI has enough polish you know side chats all these things that like Codex app I think is really sort of pioneered as far as like good like agentic development environment experiences um you know we're starting to try and port that over so we can daily drive >> uh Yeah, I mean there's a lot of nuance there and I think like on the point of pairing, I think there's this whole spectrum of how do I work with agents like there's a whole co-pilot model, you know, back in the ancient days of what like 2022 where you use uh AI is just like autocomplete to kind of like this agentic teammate where you know you you tell the agent, hey, do this thing and then you know you expect sometime later there's a PR that's merged that's tested and already working and then like in between is you know you're using the codeex app or you're using the control UI to actually more directly like work with the agent on like a turnbyturn basis or a loop byloop basis or you know whatever your favorite kind of um uh meme is of the day that >> um could you somebody asked what is uh what is the point where agent agents LLMs are failing for you >> um Asian where Asians are failing. Um yeah, I think this is one of those things where it's just like it's always incredible how easy it is to take something for granted and normalize something. Um I think I used to work at Amazon like Jeff Basil called this like divine discontent of the customer in in the sense of like you will never be able to satisfy what customers want because the moment you satisfy something they will want something more and >> being married.

[00:20:03]>> Oh sorry [laughter] >> I think there's many metaphor yeah many things this could apply to. Um >> and the same thing with frontier models is just like you know it's amazing what these models are capable of. Like if we want to talk about super intelligence, you could make the argument that we're already there. Like if you talk to like a 5.6 or a Fable, like they're probably I mean like they're smarter than me in many ways and probably most people that I know. Um so I just like start off with a disclaimer that like holy crap like these models are incredible. we're living in the future and yet somehow like I think it's like you know I uh we keep finding like oh here are the things that uh they don't do well like for example figuring out where the car wash is. Um okay and I guess like that aside right now where I still find struggle with the models are that they're good at like let's say coding you know like >> I think there's this um driving narrative that you know what is the future of software engineering it seems like uh these models can basically do just about anything right now and like TBD I think like that is still like the trajectory where we're headed but I would argue that at least currently if you look at 5.6 were even fable. Um, you still do need human taste and human judgment, especially for bigger projects. Like these models are generally good at figuring like writing code that works and if you give it acceptance criteria of like what the behavior should be like making sure that the functionality is actually there but they're less good at okay how do we make sure that this code is maintainable over time and like the test I think about is just like in 6 months from now if I like keep iterating on this codebase with agents like is that codebase still going to be legible to like either humans or agents um and And like common failure modes I see for all these models is just hey for example if you want to like enforce a security invariant like make sure that um oh every time you you know make a call make a network request maybe uh like check your security policies and like um more more often or not like the model today would just like add if and else statements across um you know thousands of like different call sites versus you know maybe there's a much cleaner a way of enforcing this in terms of like defining like a software contract of like hey there should be a service that makes all network requests and then you can like centralize all the logic there. Um this is just like an ad hoc example but generally like there is like the if there is a way of just consolidating um and doing like good coding architecture for like more complicated services. um the models will generally just choose writing more and more code which works in the beginning but then quickly becomes unmaintainable and right now like this is why like human judgment and human taste is so important. Um, so yeah, ironically, uh, I think like coding is still my biggest complaint of like I I don't find models to, um, I think there's still a ways to go before, you know, you can completely take your hands off the keypad or, you know, your voice off the whisper flow or whatever it is that you're using nowadays.

[00:23:31]>> Yeah.

[00:23:31]>> Yeah. I I feel like for me the number excellent uh overview my like number one annoyance is just like jargon. I feel like the past couple days like open cloud team it's like >> all of the like AI generated PRs we get it's like I don't know how many times I have to read read the word like seam or loadbearing or like >> like my brain just fries when I try to try to read a lot of these PRs. So, um, sometimes just like knowing how to like speak the same language as the model is is part of the problem. But, um, I know we only got a couple more minutes here.

[00:24:05]Uh, Kevin, I think maybe like last question for you is what's like number one thing you'd like to improve in in OpenClaw.

[00:24:14]>> Number one thing, um, so so many things.

[00:24:18]Um, I mean I like the self-s serving part of it would just be like, you know, make it enterprise ready. That's uh essentially what my team's working on uh you know with the OpenClaw Foundation and everyone else. Um I think otherwise I think for Open Claw um I think the one thing that we could do a better job at is just making it easy for people to get started. I um I think this is something that like you give people an Asian today um and it's just like what do I do with it?

[00:24:53]It's like anything you want. And while that's true, it's also hard to kind of action on top of that. Um, so, you know, I don't think this is anything that's like technically sophisticated, but just packaging really matters and the first time experience really matters. I think um you know openclaw would um could be used by a lot more people if we just focused more a lot more on like what is like how what is some opinionated ways of how do people get charted and what are the best ways of using it.

[00:25:30]>> Yeah. Uh to that tune Patrick we had talked a little bit about this beforehand. Um, our current version 7.1 is um, in need of some love, some fixes which are definitely paired into 7.2 coming. But what is the real big what are some of the big ticket items changing in 7.2? Do you want to cover those ones off top your head before we close her down?

[00:25:57]>> Yeah. Yeah, let's do it. I'll just kind of >> And when are we releasing it? Maybe.

[00:26:00]>> Yeah. Yeah. Yeah. Yeah. We're on beta 4 right now, I think. So, I believe we're targeting a release date like this week, like ideally like next couple days.

[00:26:08]Don't quote me on that, but uh I think release is cooking and it's in progress.

[00:26:12]The big things in this release are the control UI updates that I mentioned. So, Peter has been cooking maniacally on making control UI really awesome for daily driving. There's some also kind of cool experimental stuff that we're shipping in there. For example, a like pages concept. So your agent can generate widgets which will also include um I'm not certain if it's in this release but eventually like MCP apps as well. So you could have like Excaladraw contribute a widget that gets pinned to a persistent page. So if you kind of want to build like a mental model of like you like Matt PCO has this cool like learn skill that like kind of develops your understanding of like a project as you work on it. Think of that. But like that plus like Tariq HTML if they you know made something super cool in open cloud native. So a lot of cool stuff coming in control UI. Um our onboarding has improved a lot. So Jason's on the call somewhere here I think. But uh we've had folks doing a lot of work on making onboarding really awesome and more like agentic. So it does some cool stuff like detect API keys automatically and wire those models or harnesses for you. And if you have nothing on like a fresh VPS for example, it can pull down like a local model and run, you know, something like Gemma.

[00:27:24]>> Yeah, Gemma 4. Like it can run, you know, some local model to help you onboard and just kind of immediately get started without a lot of the configuration headaches. So to a bit of your point, Kevin, like trying to make it like easier to get from zero to one with your first claw. Um, on some of the more like administrative side, like we've been doing this like massive refactor to SQL light. So, previously we had these like static JSON L files that when you run a claw for 3 months and you have like a 200 megabit file is just going to blow up. Um, so we've been cooking on like a large refactor to SQL light that'll also enable a lot of kind of cool portability things in the future. So, um, some cool work there.

[00:28:06]Uh, >> what about LTS? I know that's not part of this new one, but uh, what what's going on with that? I see it cooking.

[00:28:14]When's it going to come out? What What's it What's its promise?

[00:28:19]>> I can actually speak to that and I'm glad that you called that out.

[00:28:23]>> Um, so for the longest time, especially for companies, excuse me, um, like stability of OpenClaw or just yeah, general stability was a big concern and so people have been or big uh, enterprises have been calling for like a LTS version of OpenClaw. Um so this is something that this is some work that I started and Don who is you know one of the foundation member has uh kind of carried through the finish line is uh what we're actually going to call instead of LTS it's initially going to be called the stable release um of openclaw and this is going to be a monthly um release where um essentially once we release it uh for a given month like the only things that will go into the release subsequently are going to be uh like security fixes and reliability issu fixes. Um and it's a version where we focus on you know supporting it for the given month uh where it's rolled out and just making sure that the installation and upgrade process is rock solid. Um that being said, I would say like open cloud in general is much more stable now and so you know you should be fine just like upgrading um to the daily version but uh the stable release is going to be for essentially companies who just want longer lead times between upgrades and support of core features. And I would say like uh we probably we will announce uh something we'll announce the details um if not end of this week then next week.

[00:30:04]>> Yeah. Cool. Um yeah, I'm I'm very very excited for that. Uh because I uh you said we've become a lot more stable and I agree. Uh with exception of upgrading which seems to break uh enough peoples that that's all I get to talk about on social media with people it seems.

[00:30:24]So I'm excited to see 7.2 um squash that and the stable release also um take care of that. So, thank you very much for joining us today, both of you. And thank you community for joining us. And I look forward to seeing you again next week, same time, 11:30 uh Pacific time.

[00:30:44]Okay.

[00:30:45]>> Thanks everyone.