Speaker-label warning. This transcript has no per-speaker labels. The talk is primarily Robert Overweg speaking, bracketed by a host introduction at the start and an unidentified Q&A at the end. Audience questioners are not named. The source also contains noticeable speech-to-text artifacts that have been preserved verbatim:

• "oak claw", "open cross", "open gloss", "open clock", "open claw" — all appear to refer to OpenClaw.
• "gbrain", "g brain" — GBrain (attributed to "Gary Tan. Of Y Combinator").
• "OB" — an open-source ambient recorder Robert has on order.
• "decent skills" — likely "Anthropic skills".
• "Granola" — the meeting-transcription tool.
• "NTP" — likely "MCP" given context ("it can connect through NTP or API").
• "60 place" — likely "feedback loop" / unclear; preserved as-is.
• "chat windows I'd like to keep it as much as possible on like our own our own stack" — the antecedent appears to be Robert saying he doesn't want knowledge to sit in vendors' / "other people's chat windows".
• "So we have developed this yet" — context strongly suggests this should be "haven't developed this yet".

Do not silently correct these in quotes — flag with [sic] or a bracketed clarification if needed.

§1 Host introduction

Hi. Everyone. Welcome. This is the last session of the day in here. You guys have done an incredible job making it this far, yay all of us. Wow amazing. Cool. This is Robert Overweg. He is going to be talking to us about something I am really excited about is one brain concept. I was chatting just a second ago because of Tessl we've been focused on the kind of linear issue to merge PR orchestration pipeline. But the more we do that the more we realize that actually pre-issue is really what matters that decision making and that scoping is actually the hardest part and kind of makes the rest look trivial. So I'm really excited for this let's give him a round of applause.

§2 Why one brain — the frustration

For having me. And thank you for staying around and I'm also aware that I'm the loss be standing between you and parking there so I'm trying to keep short but also keep it interactive. Today we're going to talk about one brain how you can create one brain for the organization I do have to say we aren't like this large enterprise or still like a startup so for us maybe it's easier but it's also easier for us to experiment and try out new things and I want to share some of our learnings and also yeah some of our failures in a way. Because for us it originated from this type of frustrations are often like hey where's this file where's this presentation you're always looking for these kind of things changing people it's really irritating at least for me or when I am working in the morning like five o'clock or late in the evening and I actually want to get a prototype out really quickly then it would be useful for me to know if they're prototype is actually. For 500 SUNIQ units or if it's for 5,000. And when that's not properly communicated to me it's difficult for me to set everything up but if I am connected to the sort of all of the all the knowledge of the organization that was also part of. Like meeting to conversations then I'm just really easily able to sort of channel that. And bring us forward in using my work. Now I would like to press the next slide ideally. And for a very small subject on hey how can we actually improve our agenda pipeline? Really really large I understand but these are kind of questions that also surface in. Organizations

§3 Who Leapfrog is

now I do have to say that we were mainly in fashion for this type of brand so big brands and we do a lot of high volume image visual work for those guys. But I also have to say we're a small team so that combined large volume work and small T and then creating these type of works what you see here is a lady that we created this digital person she is currently being hired to or rented to so he'll figure out the client and G star. And we also create this type of imagery video based through AI we do that at scale so we do these type of things. But we don't just do that one thing we have got five six more modules so we need to produce a ton of content a ton of stuff so we really can't be occupied with trying to find folders trying to find stuff so we really have to work on AI native way just to be able to cope with the gigantic output. So that's basically the sort of setting that we're in and I think that's really helpful also for for innovation because we can't have to.

§4 Two kinds of knowledge

Now Ms what I said at the top so we're trying to continue to learn about everything and we're also continually trying to service the right information at the right time. We see two types of information one is company knowledge so things like new developments new things are happening in space in terms of like go let go of skills research Visa create making medical knowledge accessible and then there's also in the creation pipeline but I'm sure that the e-balls are stricter making sure the skills are getting optimized so there's two kinds of knowledge that we try to optimize for.

§5 The starter stack

Now we started really simple and you can start in such a way as well. So we just started with oak claw on the sandbox environment that GitHub is repository that has the data like all the research those kind of things and then we use I use obsidian to also have it locally on my system so that I can browse through it and I'll show you how that's set up as well. I find it way better nostalgic notion always was so slow to load and really clunky the green blankets or people are fan but I'm just more of a fan of city because really fast and snappy and then telegram is yeah the surface through which I can just communicate 24/7 like a little bit of subtle because you know you could wake up in the morning and then you can sort of ask your agents or I walk on the streets and then just talking into my phone like this to find information or to sharing information. Really really useful now we are actually at this stage and I'll show later how all of this works it can become it's not necessarily a complexity it can become more complex gradually but it's good to see that some people take photos but it's good to start here I think because then from there you can just continue building yourself but I'll explain these things a bit later and also what are specific setup was for each of these components.

§6 What you get back — natural-language sparring partner

Now what you get in return like I said before I have sort of all not all most of our company knowledge on my phone and I can just send all the message messages or type myself the surface information which is really really valuable. So one example here I can talk through natural language so I again forgot what we wanted in implementing our CD pipeline but I did remember was from Microsoft and had like these four steps or something so just asked my agent what was it being again for the four CICD things? But then it says oh you probably mean this and this and this and this which was actually accurate and then it also because it understands my context it also asks is this for the talk which is also really useful so I just continuously have this sparring body so what we're doing is we no longer search for files and more like for ideas and context and those kind of things. Which is a massive game changer.

§7 The research agent (cron job)

Then we have agents as well so I have an agent. That proactively does search for me and I've agent that is able to interpret search and I find myself want to show some examples here as well maybe you guys have this already right and this is really also quite basic but we have a crown job I have a cron job that runs on my open clock that tracks certain accounts on hex those accounts just on just some specific words so you can think of a genetic engineering or any words that relate to such pipelines that I'm interested in I follow some niche accounts as well or niche I think like 120k forwards maybe not that niche but I do say in touch with everything that is happening and it's almost like my newspaper because I don't read the newspaper but I read everything evolving around agents, the eye, those kind of things. And that's the service to me daily in the morning.

§8 Promote-to-vault discipline

But what we did figure out what was really useful is that obviously you don't want to put all those things in your own vault or you know repository. So what we do is I promote stuff to the vault which is really important you need to remember this. Once after a while I figure out that it actually really important and the great meetings out of it and those leaks there surface to the wider teeth because I don't want to bother everyone with my unvalidated research. It needs to find some grounding in reality. Okay. We visual here is also a very good example. So I found this skill from a decent skills part of a larger scale patch from adios money guy from Google. This was for. Skill breakdown larger tasks. I thought okay this could be really interesting but here I just asked my agent like hey is this actually interesting for us and since it understands which skills we have our cockpit setup what we want to do is able to analyze, I need to steer it a little bit because I need to say don't we have surgeries thoughts breakdown already it's a developer we work with. And then it's like actually I'll have a better look then at the end it's able to say summarize because many skills have nothing you don't already have your sbdv task requires actually more detail like hey also if you look at the time here it's 5 plus 8 here and it's 6 plus 8 at the end so 1 minute you analyze that scale came to a conclusion and we can decide that we're not going to put it in the vault.

§9 Inside the vault — research + notes

Now I want to show you a couple of things. That's why I have this checklist here so I don't forget. So in our vault I have a lot of research as you can see here on the left. And that's our router structure it's almost 1200 files this setup is able to cope with all these things without doing additional. Stuff we started doing additional things like that kind of stuff but it's not really needed. In those files. I add my to do's also related nodes sometimes I create connections myself and like actually this file needs to be connected to that so that we keep it as sort of like memory hub together. And also if you look if you look here so here I asked my chat agent so here in general how did open cross search work again then also like hey can you clarify what you mean are you thinking of xyz and then I clarify and here it gives me exactly how it works. So yeah I can also even keep up to date with the things I make of myself. And then here we have like the daily update here I also get sets of dates on security things here you can also see that new meetings have been had by other people so I'm just getting notified like hey this person had a meeting that transcript is available for you. So I've shown you the research and also the notifications. Also if there are any questions. Yeah, I guess we have to part ways a little bit too and.

§10 The second vault — client information

Then we have another vault which has all sorts of client information so we work with like I said earlier a lot of files and also a lot of our clients work in a variety of different ways so some work with my role to out ours some work with tarpaults some for keynote some work with figma and it's quite difficult to get people to change and work in different ways so we sort of figured out let's just interpret whatever they give to us and then put that into our system. And that works really well and this entire system is also based on natural language so you can just say hey what were lost files from Calvin Klein that we got delivered from ex which is really useful when you run production so whereas an asset that's no longer an issue which makes me happy because you can then just focus on building things. And we need to build a lot.

§11 Dorsey quote + flatter orgs

I think this is also part of. A sort of larger fundamental philosophical concept and this is from Jack dorsey founder of Twitter and dogs and he said the emergence of AI is generous and premise of hierarchy as a requirement for the first line so where we have you know in the enterprise business all of these levels with middle management which we sort of used for communication but there's also a lot of issues there. So what if information can be modeled on stood and distributed in real time the organization still needs coordination mechanisms centered on people what if because agents are sort of smart enough to keep interpreting everything not always sometimes they interpret it in the wrong way but mostly it's really useful. So we thought. Yeah should we just make as much as possible humanly no pun intended available for 40 AI so that's every conversation that we have meetings research also code base obviously.

§12 Record everything (Bridgewater, OB, Granola)

And it was also a little bit inspired by what I read I think almost almost 10 15 years ago. By ray daily had this hash fund called. Bridgewater yeah but they recorded every meeting back then already back then I thought was yeah a bit weird but maybe I was just in a sort of a low trust environment because right now I feel like you should record everything because everything that is not like digitized sort of doesn't exist and you're already a disadvantage if you don't do that. So now we also order this thing still waiting for it's called OB is open source application open source technology which records what we say and well when you find this antidote because we're waiting for it for like three weeks but when you have like coffee outside or somewhere else you can just put it on a table it records it and then you can send it to your own servers so there's no like subscription stuff which is really own the data which you think is really good. And then Granola is a very useful tool for us as well. Which transcribes meetings but in a proper way and you can also type your own notes at the same time and then when it makes the full transcript it prioritize molds it around your notes which is really useful so here we are in the view that I see my co-founders all of is me notes which is pretty useful and it can connect through NTP or API to it. Which is also nice. And then afterwards when you have like a clock design or an open design if you have those read nodes already you could practically generate the presentation that you then need to send to the person that you talk to semi automatically at least you could get like first draft. Which is also. Pretty nice. So these type of questions that we have here we can all surface this through the meetings that have in your core that so I think if you work more remotely and more digital. You could just record everything it's really beneficial so it could be more beneficial if you are further apart which is kind of funny.

§13 Chief-of-staff agent (planned)

This is our to do. So we have developed this yet but as a co-founder like I am in sort of production I'm in sales I mean in positioning we're trying to do everything at the same time sometimes you maybe forget to off like the right type of sales questions so when we have meetings but like the things that we need to ask in these sessions we do know so what if our chief of staff just surfaces these type of questions that we need to ask to the client in that specific call so it can also help us think along.

§14 Own your stack

This is maybe a useful message for you guys like I would not want my knowledge to sit in just a lot or in other people. S chat windows I'd like to keep it as much as possible on like our own our own stack and our own service so we really query it.

§15 Keeping it real — failure modes

And the keeping it real section because it's very very messy to. Build stuff and so this was for me and for us. So we had a lot of conflicts and loading issues you can update stuff on the CDN as well and then you can have like all sorts of merge conflicts with your actual github repository which is not fun. This is an interesting one knowledge sharing like in general is could be rather difficult because are you really going to share all the meats that you have with everyone? Because what if you mentioned something that really should not leak out? So we have certain setups basically for that make that less of an issue. And yeah so we made some boundaries between people teams and clients and just knowledge moves throughout the stack another thing to call out that as you guys already know as well right the AI can be brilliant and dumb at the same time it can sort of lose sight of the most obvious things that you thought I'm like I said it's already 10 times or it just doesn't understand like very basic things but can be brilliant at the same time. And then I had a lot of issues with just script breaking and chronic jobs not working and that kind of stuff. So it's pretty yeah once you got it set up it works but then when you want to start adding new stuff to it comes to full-time job.

§16 The actual architecture

So to come back to that early complex role that I showed how everything was connected. This is how it works for us so I've got my own private vault with all of those research documents and that kind of stuff that's connected to github I did add gbrain to it that's from Gary Tan. Of Y Combinator and through that I added a pt factor and this is also makes use of zero entropy so I've got vector keywords translation key pages etc. And but I do experience that I don't really need it for the amount of files that I currently have it seems to be that like after a larger amount it becomes more more useful and then we also have a chief of staff which is also on a separate instance. Because I don't want everything to sort of leak over into each other and then telegram is a way to at least for me educate. And like I showed with obsidian and looking into neo4j to like add another layer of knowledge. Sort of filtering to it.

§17 How OpenClaw search works for them

And then here's an air dive. Actually how the open cross search works so and we have memory search so semantic search over memory.b and top of files serves as preferences decisions and past contacts and we have the vault context loads of relevant slices of the vault by TOS time columns a vault that's our enterprise stack for self-serve. And then dbrain if it's like a large research and then we also have direct file reads. And for example for today the top questions hit the death consoles directly some stuff on haiku because I want to reduce our spend because can be a rather expensive to keep this running especially if you run it for multiple people.

§18 Advice for getting started

So my advice would be to start small. Let one person suffer through first don't roll this out to everyone. And yeah once people like really really local just on you can also just run it on your own laptop to make sure everything is working properly and make sure that you have proper security hardening and everything that you do and make sure that you also align with what you're actually allowed to do with the data. And I would also set up a chief of staff agents because people often say that they don't have any time which is true because they are so often in all sorts of meetings and finding stuff but if you can give them an agent that sort of shaves off 20% of all of those jobs and they find a little bit of a nuisance that they get a time back then they can actually move more into these type of optimization things. And then the research agent this is just magic. To constantly get that information and be able to learn from the world through your curated lens. And then yeah basically see what you want to promote to the wider group. And that's it from sort of how we built our own internal brain and our almost like our eyesight and if there are any questions please feel free to ask.

§19 Q&A — scaling client files

So interesting I honestly think this kind of thing is going to be like table stakes for everyone in tech in like a year's time. Does it even have any questions for Robert?

Hi what concerns do you have about like growing that customer files amount like how do you see that scaling do you see degree being able to handle that or do you think that you're going to have to go near 4j or do you think that's eating something up for that?

That's a good question. The brain that works with the company files that is still fully local one person's laptop because we're still figuring out how we want to expose that so unfortunately I don't have an answer yet to the question like locally it's able to surface all the things that it needs but we still need to look into stuff like permissions and making sure all the data keeps being segregated like its own buckets yeah so you build a different data types as well it's going to be a real struggle. Correctly.

§20 Q&A — time cost to set up

Again how much of a time cost was this to get everyone set up a long time ago to maintain.

The bottom of my mind was in insane amount. But also it's a little bit of like a hobby. It's become a bit obsessed and so in my DNA that I just continuously want to learn and really say so it's just part of my nature so that's reseal my work. But. Difficult thing maybe like a month of work to get these things set up but the good news is once even open gloss I get up with the. Right container soccer with the right permissions can be an issue but once you've got it set up it's also script and it can be 30 minutes or 30 seconds. 30 seconds. Turn up yeah once you got this set up right once someone suffered through it other than the update configs and then they come with a new version and I had the other day like half my things broke it's not that deal.

§21 Q&A — team interactions; AD preferences

Hi. How are other team members interacting within subject source of information for your team both said so difficult but mission learning and expensive and how how far from other team members interaction with this so I had another kind of different question though is also. Just I've done some flat work in actually industry as well but actually understanding how this particular quality of information is like you're just allowing the AI to kind of structure itself and understand semantically what the information we're working with or you're thinking like you know infrastructure about how to optise this kind of information as well.

Let me first try to answer the first bit how people interact with it so like 80% of the stuff you see here we keep with just with the leadership team at first and then we promoted with all sorts of knowledge to the wider group and that's way way less that's maybe like 10 to 15 things but they're rather back so that could be how to work better with a gentle how to reduce bugs how to do those kind of things and to answer your question in regards to the client data you have now over I guess all familiar data points on what the art directors like so there's a lot of subjective and yeah we just interpret that we do have a 60 place that also feeds that back into our production.

§22 Q&A — why OpenClaw

Where do you pick open glory as your orchestra and how have you found it suitable or unsuitable for this exercise?

Yeah, so I found it really suitable because I don't need to have any other like actually I don't need any other stuff like gbrain to in-depth data when it's almost like 1500 different rather large markdown files so it worked really well but I didn't do a very large industry wide scan of everything that was available I think I started using. This rather quickly after it sort of got got released and yeah just really like it I like it's proactiveness as well.

And I guess like you're now able to build more agents on top of this within within the same orchestrator as many as you need on top of this knowledge correct yeah yeah yeah the place that the V's get promoted so that is on like a separate database and then other agents can also talk to you because they don't want that on my open claw setup I do want to keep those. Things separated.

§23 Q&A — Hermes and coding

Thank you for talk. I was going to ask actually a similar question have you thought of using harmless because that's why I decided to use and also when you ever tried any coding along with the agent trying to. Streamline coding processes but or is it just about

no definitely so first on Hermes so they came out after club. So I did wonder if I should set up a different like an additional agent to troubleshoot all the things that consistently break so I thought maybe use hermes and then find out how that works but it just didn't have sort of a mental bandwidth at the time to look into that. But since this sort of works I don't really have a reason to look into that. What was your other question sorry.

I also tried to streamline coding yeah so we're working on a full dark factory pipeline and open cloud is the orchestrated there which triggers the right skills that other industrial agents do reviews and those kind of things. But it also just connects to codecs and that does the bulk of the work but we do own the orchestration and all of the. Basically each harnessing kind of stuff around it but that's also a constant I want to say metal but it's a constant let's say experiment.

§24 Wrap and logistics

Okay would you guys mind staying for 20 seconds? After we give Robert a big round of applause thank you very much. Hopefully Robert will be at the party which starts in now so you can ask him questions over a drink if you fancy it. Two quick things to say one the draw for prizes including those meta ray bans that I know you guys want will be tomorrow evening at the end of the talks the other thing is there is a tube strike tomorrow. Sorry but it's still going to be super possible to get here Farrington's nearby there's a Melee bus around here in the middle of London you know too cool let's go party thanks everyone. Yes it does yeah.