AI Working Group, September 2 2026

Hello, Reuben. >> Hello. >> How is your family doing? >> Yeah, you're fine. >> Good. >> Getting older. All these children and me. >> Yeah. >> The thing that happens automatically, I guess. >> Yes. >> [laughter] >> as as I've been putting in my writeups. >> I'm a seven centennial man. So [laughter] >> obvious. >> Hello. right by hand. >> Let's see how many people join the call. I'm pretty sure Wayne was planning to join. It's on his calendar or at least this window is blocked off for it. >> Mhm. It's a little bit vacation uh season. >> Yeah. >> No, it's all it's done. >> No, it's not locked down. I I'm leaving for four four or five days starting tomorrow. >> But it's fall like the leaves are turning brown due to drought. But [laughter] >> well, in California, the leaves are, you know, we only have two seasons. >> Oh, yeah. >> Green or dirt. >> Yeah. They uh there's like one or two blocks that I know in this area that actually have trees that actually change color and actually fall looking at stuff. So, and then then I'm able to see uh sometimes snow on the mountains, but that's that's about as much snow as I get in this area. [snorts] But >> at least I can drive to it, which is a good thing. There [clears throat] he is. >> There's Wayne. So, when is Stratles going to support all this agentic stuff? >> Stratos. >> Tell me. You tell me what you want. [laughter] >> Just AI everywhere. I I have no problem putting AI in as long as you tell me what what do you want >> I mean right now >> a dash of AI maybe like start with a logo that's >> with a plus AI and then or LLM powered or something >> they should just dance basically right like one of those little dancing logos >> yeah and I want the UI to be different every day right like just AI [snorts] generated so if I click an app that it's like just [laughter] Well, >> takes into account my mood. >> Yeah, you already know I have two personas, right? I have my AI persona >> and I have my, you know, my human persona and I always argue and approve, right? [laughter] >> Flip brain used to be a networking problem. Not anymore. >> No, daily reality. Actually, there's a lot of AI in the only way I could get done the work I did on Stratos was with the you know AI work because I literally >> I noticed >> Yeah. >> Well, I'm not hiding it. >> No, no, I mean no like my my RFC that I did uh or route identity. No, what is it? Route policies. >> Mhm. >> That was fully done using AI. I have not written a single line of code. >> It's got qualitywise a lot better. I mean, you still got to like watch what it's doing carefully, but >> I find that like acceptance testing is more important, right? So, creating an environment where it can like iterate and have a feedback cycle and then you get good results. >> Yeah. Yeah. I got the same same thing. I mean, that's the type of thing that I think we should be able to somehow incorporate into into Cloud Foundry is some sort of loop, right? >> Yeah. Like ideally you can like what I find really valuable in my home lab is having like a borch director and like a cloud foundry because I'm working on cloud foundry so that my my ed uh agent can iterate like do actual borch deploys of the source code or of the changes and then debug the components right so you would want a similar thing to happen for apps right so that I don't know if that should be the initial thing but ideally maybe eventually you can have your agents run on cloud foundry so that they can iterate with uh on apps. >> I think that's a great idea. That would be great. >> Yeah. The the thing I did for the my stratos talk is I wrote it from my AI persona point of view. >> So when I said the AI way I'm talking as if through the AI saying you know you rejected this but we did a lot of looping. The reason why he approved so much is because we did a lot of uh approvals up front so I don't have to worry about it. >> You know, uh when Chris put in his couple things that he wanted, I went back and forth a couple times with him before we do approval. So, it's not that it's not approving, it's just a different cycle. So, uh, I have a little picture of plan, argue, execute, um, redo because the plan was wronged and start from the beginning again. And there's that that was kind of >> the picture I put in place. So, that's how I'm doing that's how the AI talk uh, sorry, the Stratos talk is set up to be uh, AI based. Hopefully, uh, Dr. go and I can work that out and make it work. Right. >> Um, should we wait for other people or do we think this is it? People accept us. I don't know if we can actually see. >> Much much smaller group than the last time. >> Yeah. I would just say get started. It's already seven after. Hi Chris. >> Yeah. >> Wild Chris appears. Hey everybody, how's it going? >> All right. How you doing? >> I'm good. It's uh you know, Wednesday morning, getting stuff done. Looking forward to H Highleberg in a couple weeks. >> Oh yeah. >> Hey, I just emailed AWS. >> Thanks. Thank you, Chrissy. Okay. Do you had something in mind, Don, for this meeting? >> I I was thinking that uh the agenda for the meeting should be to figure out how to organize the followup meetings and the working on the ideas and the uh uh the on the ideas. So how to organize the work and then uh look into how the ideas are clustered in uh um so um what kind of ideas we have and how we want to approach this maybe uh one cluster per uh meeting or something like that. Um yeah. >> So basically figure out how to uh >> um tackle the ideas and and the research notes so that we at the end come up come up with um proposal what makes sense um for the different uh clusters which idea we think is the best and whether it makes sense to be integrated in cloud foundry or not. Um >> thank you. Um, I did a bit of that. >> Um, not the meeting thing, but I did have AI do a bit of uh rating the different research things on different axes and um that's fun, looks nice. Um but I also did ask to um figure out what the uh yeah what what strategic idea or clust potential clusters could be or areas of interest and it found three of them >> and I think that those are also things that I had like found a bit right so it's the durable execution. >> Um, so that's so that's about the the the checkpointing and making sure that these things can like it's a different type of workload, right? It's close to a task, but it has a bit of state in it. >> Yeah. >> Uh, and then there's the the snapshotting and like how do you want to deal with, right? Like on one end we have that RFC that's in draft for like actually having volumes with state in containers. On the other side you have things like dapper and those types of things that have more like provide APIs to do snapshotting or um checkpointing. >> Yeah. >> Uh something along those along those lines. Uh there's like stuff around uh identity. Um there's a bunch of stuff in that area. And then there's the the isolation part like an actual sandbox, right? like um a sandbox API being able to for a given session spin up a container that also has some sort of I mean there's a bit of overlap with like the durable addressable execution and the the isolation aspect of it but uh yeah I think those are distinct s uh problems um >> yeah I think one uh sandbox this isolation sandbox box is uh something which uh people discuss a lot in the sandbox. I also have in um scope like uh controlling the out egress uh communication for example and access to secrets for uh >> yes >> yeah and auditing that right >> and the and the one I added as well the identity portion which is always obnoxious but necessary. >> Yeah. Uh what are the what are the diamonds on your little graph here? Is it like >> Oh, so those are ids. So we have ideas and research things. So there were some things show put in as ideas. So uh so we can go to the full markdown. That's this one. This is this is by a notable author. >> No, no, no. Ignore that guy. [laughter] He >> does not know what he talking about. No, but so like those are Yeah. >> then. Yeah. I mean that when I was reading through the ones you were doing, that just seemed to be the only gap area that I could find a like a missing that you weren't kind of covering. Um not not not no statement there other than like that that that's the one gap that I could find when I was reviewing. So I tried I tried to write to that gap. >> Yeah, makes sense. Then there were some research notes already for this which were related to there were some standards right around this. >> Yeah, I think I linked to them in the footer. >> Yeah. Okay. Um >> what's the other triangle? >> Um this one. Yeah. So that was one. So I've looked at like the um dapper seems like an interesting technology uh and we already have side uh sidec cars right so this could be potentially a thing that putting sidecar >> what is its like adoption in the ecosystem is like >> uh it is um what's it >> like it's not something that's going to possibly just disappear like >> it's CNCF have uh and it is has been um what's this? [snorts] How do you call that? >> Graduated. >> Graduated. Yeah. >> Okay. >> Right. So that's why I thought that made it a potential interesting thing to put in. Right. >> Okay. Yeah. >> That and it has like a lot of client app clients support type of things, right? So So that was one of the So we have a few ideas, right? So there's the deapper stuff [snorts] and then you could even make router the go router deapper aware right so that it knows for example uh where a certain session is being run right and like it has checkpointing so if uh a container gets killed it can uh reassign that work to another container and then start the work again from the last checkpoint and then the go router because it is aware of the the the chain of like it has some sort of hashing thing. How does I don't know what's called again? Um but yeah like it it keeps track of where things are being executed right so then go router could be aware potentially >> customizable routing algorithm h based >> yeah um but then I was thinking about this dapper thing and then I thought maybe we just want durable tasks for CF that was my like thinking along this >> I mean that's actually where my mind was going with the Dapper if we can like kind of add the missing things. Uh >> yeah, like maybe you don't need all dapper, right? Yeah. >> And then just oh look to use dapper you just configure this or whatever. >> Yeah. And then I even went as far as I don't know where I did that. Uh I had another idea which was basically durable tasks maybe even was that the one um because in the end I realized like okay if we're we already have staging which is basically a process that runs and at the end it snapshots its state [snorts] and puts it in a blob store. Right? That's basically what we're doing when we're creating a droplet. So that concept already exists. So couldn't we do the same for agent tasks, right? They run, but then instead of running to completion and doing nothing, it would snapshot its droplets. Not not the droplet, but its workspace. Maybe it has a workspace directory that just gets snapshotted. And >> the back end of things uh things like cloud code, they have this space, they have called the scratch pad. So >> yep, 100%. >> Yeah. So basically what you end up with if you want to run agents on cloud foundry you basically have the getting your runtime stuff installed which is basically staging and build packs right so you get your droplets and then you have your runtime state that you may be also want to snapshot so I thought these concepts are actually quite close to what we already have in cloud foundaries just tying them together a bit >> what I'm what I'm wondering is that we would have like agentic build packs we would add like so we'd have like a cloud code build back a >> um maybe we could use PO or whatever makes sense. >> Yeah. And and you could have an MCP server where you ask for um where an agent could ask for a different dependency to be installed, right? So then it would be the container would go through a staging cycle again. container gets killed but the the state is snapshotted and then it's brought back resumed with new with a new droplet and the same works workspace. >> Yeah. So I wonder if we can map MCP right onto the OSB API and >> I don't I don't know. I'm I'm thinking out loud, but um >> yeah, I mean >> like I I think it would make more sense to maybe have I'm lost in my Windows. Maybe have uh cloud controller have some sort of MCP API so that you can have like agent identity. You can use maybe MTLS something using the identity search that you already have in your containers to talk to cloud controller to be able to control the life cycle of your own container, right? So you can ask for a new revision with new dependencies or you can ask for a snapshot to be taken or whatever, right? Like that's that makes sense. And the reason I was suggesting the OSB API path is I was thinking like I was viewing the MCP connections and as sessions as like on demand service creations kind of thing is what I was thinking >> and then so we use what's already there and then on the build pack side uh where it's going through the staging and stuff have it read from the vcap put it into wherever it's got to go for configuring to use the the service instance MCP service inst instance that it was provisioned. That's all I was thinking. Not not >> um >> yeah, that makes sense. I mean, there's still like the use case for more MCP servers, right? So, you you probably when you're doing this agentic stuff, you probably have corporate MCP servers that provide access to some unique resource that that that's specific to your organization. If we had if we had like an MCP broker where they could just kind of register their custom MCP with it and then it facilitates via this mechanism then the agents themselves can say hey I need one of these an instance of this then or I need >> yeah and then the agent build pack could be aware of that right and inject it into it right yeah >> yeah so that's what I was like if you play around with these building blocks that we have in CF a lot of stuff maps pretty cleanly >> close very close yeah >> yeah it's It's really close. So the big question is where to start? >> I think uh I think we should we have the clusters to my understanding. We have clusters like durable execution. You mentioned sandboxing isolation >> um um framework SDKs I don't know uh they are maybe more into the build pack space. how to expose MCP ser >> and uh every workshop or uh we can build working groups which are focusing on the different clusters and they try to somehow propose uh solution for that uh cluster which technology or what makes sense actually uh uh I liked your uh uh diagrams at the beginning I couldn't get whether they per quest. If we have something per uh we see >> it's currently being built that my agent was not done in time. [laughter] For example, if we have this per clust where we have different proposal how we which technology could be used, we can see which one is more m mature and then uh try to think through how that can be integrated whether it makes sense or not and then uh come up come up uh uh with a solution. I I al uh there are some people who are interested also from MSP but they on vacation I mentioned at the beginning. So at at least from SAP side the people on vacation >> uh three four colleagues they would would like to join but uh no >> we can have another one soon and we can we can even meet up in person soon. >> Yeah but I I need to present something the outcome [laughter] but I but I think like if we we just have these three clusters and we have all these research things I think that that is actually a good thing >> the diagrams with what what how the technology is. So yeah, I think this is quite >> the other thing is like I was thinking is like on principles uh of approach and stuff like that. So rather than just saying, hey, let's integrate Dapper in directly, I like the idea of ex extending the functionality that's already there based on what's already there, like you're saying with the durable uh >> items and the principles of just extending what's there to be able to then support that via plugins or build packs or any other mechanisms. I I think that's the the best approach. Um >> yeah, staying CF native, right? Like CF try keep keep our own spin on things but then but it's more about like learning from these these other things like what's there and what can we >> and what do we need to add to enable those use cases >> yeah and I might support >> probably there's a big chance that I will do a dapper P just to see if it like what it would look like but then more of like for learning right so you see like okay where does this how does this feel right like what are the things that feel wrong what what they encounter and take that like >> yeah when you do that PS PC I uh what the question I am here interested do we need a new life cycle for this workload or is the current one okay um so uh for example we have in the current one evacuation uh option and that evacuation could be something like persisting the uh workspace or the workbook or something and >> we have also uh hooks when the uh container is starting again and that can be used. Is this enough or do we need uh something uh more uh uh more uh uh a little bit different a new life cycle which fits better to this uh uh concepts they have in the durable? >> Yeah. So like >> Yeah. So we have that whole RFC with the discs, right? So but that feels like ah last resort kind of thing. I mean we started there. >> So it feels like >> it's it's heavy. It's it's large. It feels >> really heavy to connect containers with volumes. >> Yeah. And then putting Borch in the critical part and stuff like that. So it's like okay what's the minimal thing we can do to make this work and still be sufficient right? So that's uh >> yeah my gut feeling is you need some uh hooks to be able to do this and the decision how is with the implement >> the person who is >> yeah another question um and like I'm still trying to find the people within ravelo bankank that are trying to run agents on cloud foundry I haven't found them I still I think it's going to happen. Um, so do you all have like good user potential user visibility into this topic? Right? Because I think it's currently a bit from a just technical interest, right? So it's like, yeah, I I would like to do these things. Uh, I don't know if it's like a a validated use case, right? So that's another side like how can we validate that this is actually a use case that people are interested in. >> Mhm. >> This could be part of your session [snorts] to ask a little bit or >> Yeah. But I don't know if we have many developers in the room. >> All platform engineers [laughter] >> looking into the schedule. I think uh people with more technical background we'll join but >> I mean developers have a technical background. Don't be uh [laughter] elitist about it >> but they are not platform engineers >> right. Uh my take on it is is that it is is it was how I dealt with plugins. It was it was the same thing. It's obviously the community wanted it >> but at the same time nobody wanted to participate in it. >> And that's kind of where we're heading towards and not I say heading towards this is kind of the environment we're in a lot. I mean if if it wasn't for um Dr. down Wayne saying, "Hey, we need stratos for for our clients." I never been working on it. I would have never put the time in. >> So, it requires some sort of impetus behind the scenes to make it happen. >> Um, yeah, >> my take is is that no one's really looking at AI relative to the um governance point of view. I mean, I'm at a different level, right? I worry about the governance because all these transcripts and if you put a secret in there because you you know you're working on your company internal projects whatever you may be doing that's gets exposed all that information is exposed and in my case I'm using you know 100% claude okay um that information gets exposed and I I had to train Claude to say anytime you put secrets into the transcript warn me right now. He keeps warning me all the time now because I'm doing stuff in my lab and I don't care if the lab gets exposed and that but if I was working on a client, I want to know that right off the bat, >> right? Because I got to tell the client that, hey, I exposed one of your secrets. >> Yeah. >> Right. So I I think there's a level there of that we might be able to do some maybe it's not a physical project but the warnings behind using AI and and tying in if if if if some secret from SAP got out uh or or access to information or rebel bank getting access to accounts or at least to the data to the accounts that's really serious right And so maybe maybe there's an angle there that you could put into your talk in terms of what's going on and what what we need to do and it's something we have always have to think about when we're working on any of these projects. >> It comes down to compliance, right? In the end. >> Yes. >> And that's like comp. It's a compliance thing and as a bank we have a lot of rules and stuff that we have to comply with. But the problem is or the problem is not a problem. It's a feature for CF. we have already taken care of all that stuff, right? So if we can make sure that these agents end up or send their logs and everything through the same mechanisms that we already have in CF, then we inherit all the compliance that we have already done for CF, right? So application logs for the banking applications also go through CF, right? So that's already compliant. So that will be sufficient for agents stuff as well, right? So and then if we just use a local LLM, we are good, right? Or we we have an LLM that we have a contract with. >> The difference is is some of those um transcripts go to third parties. >> Yeah. But that's a contractual thing. >> Okay. >> From a bank perspective, right? So we just have >> So you have to have a contract with each one of these um LLM companies. >> Yeah. And so by running your work your agents on a platform that you control, right? Then you can enforce these things glo u centrally, right? Because you control as a platform you control which uh LLM are being brokered to these agents. But if you have people running all this stuff on their laptops, you don't have that compliance, right? So that's what I think the case is for running agents, actual developer agents on top of a platform because then you can do the compliance centrally. >> I think I think here are two aspects. One is uh the observability of the agent what the agent does. The second one is to uh sandboxing to not get give access at all because at the at the end it's running uh >> untrusted code. Um yes and uh you better don't provide secrets >> right >> it's it's glass radius controls to that >> right and but you do but how many how many how many of your uh developers people are running stuff on their laptops yes the laptops may be >> it's a shadow that doesn't exist at the same time it's not a thing nobody has that like I mean it's doesn't exist So uh that's why if we uh uh we didn't talk about prioritization but if we prioritize some clusters my uh uh at least my opinion is that the sandboxing should be somehow uh the first thing you look into to at least say people hey it's secure to run your agent here you can connect MTP servers you connect other backing services in a secure way and you have certain possibilities to uh implement certain use cases and then we can iterate from them. Do you >> Yeah, because without snapshotting it's already valuable, right? Because then sure you sometimes lose an agent but it can run, right? Yeah, [snorts] that that's true. And then like with the uh additional identity stuff that makes it so that you can do you use like workload identity federation, right? So you basically make it secretless, right? If you make it secretless, better, right? We can have dur ephemeral secrets and that's used to authenticate with everything else, but it's tied to one single container with one agent, right? That's great. That's another built on top of that. And then there's the being more resource efficient, which is basically the okay, we're going to do the durable addressable execution thing where we're going to do checkpointing, snapshotting, and we we can be efficient with this. >> Yeah, that's that's great. They build on top of each other. I also think that on the observability side we have uh uh certain things on CA but some looks to me like people are more on hotel uh APIs focused will be then maybe interesting to extend CF. So you uh that we offer hotel based uh uh options to uh consume the data the audit works and uh and the uh agent walks. Um there are already some uh is some support in CF and I think this will also be interesting for all work quotes to offer also autobased options. Um and that will also comply with what I observe in other areas how observability is uh >> so so we agree that an initial good like proof of concept to unlock all this other stuff would be to to have an agent running on CF [laughter] >> insecurely like we we just have one and then we see like can we make that work and then we're going to bolt on all this other stuff, right? And make it better. But that's like the initial thing we're going to try because an agent and then I'm talking specifically about something like open code, right? Like that to me is an agent because there's al this whole class of things which is like these all these agent frameworks, right? So I think in the end probably developers building agentic stuff want to use more of that. Yes. Yes, >> but I think if we can make it work for a coding agent, make it secure enough for that, then we also built the primitives that would work for for those um those frameworks, but that's like a hypothesis, I guess, or that's um I'm not sure if that if it translates. >> It seems like we should also be looking at harnesses uh especially the bigger public ones like you just mentioned open code. So maybe that build when we get to the build packs maybe the build packs are actually harness packaging. >> Yeah. >> So you have a build pack for for open code you have a build pack for >> pi.dev you got a whatever cloud code open yeah >> I also put that in that space. So in the build pack space to take care of the um >> so but I think these are distinct use cases right. So there's the we want to provide agents [snorts] a way to run agents closer to the to the app right which is running agents on site on top of a cloud foundry and I think there's a a good use case for that or there's a case to be made that it has a compliance benefit to be able to run these env uh agents in a controlled environment compared to on developers laptops. Uh I think that that that that that's at least my um my experience. Um and then there's the all these agentic frameworks for building agentic apps, right? And those should also be able to run properly on cloud foundry, but it's different and it could be that you're using an harness to build an agentic app, right? Uh but it's it it is two distinct use cases. >> Yeah, >> they are a bit different. And for example, if you're running an an agent on top of Cloud Foundry, you have things like wanting to do to be able to install dependencies, right, in your session, which is a a different problem. You don't have that problem when you're building an agentic app, right? because it maybe maybe it spins up a a sandbox. >> I mean that might be like where the durable >> sandbox >> that might be where the durable part comes. >> Yeah, because then you you might have an app, right? And it spins a sandbox and it wants a Python in that sandbox because it writes a Python script to do some user um thing, right? It wants to parse some data for a user, right? It does it in a Python in Python executes that script and it needs a Python runtime. So it has a bit of overlap but I think it makes sense to at least call out these two personas right as separate target personas. I mean there's there's even more personas actually in a working group, right? So you we didn't even touch upon the running LLMs on Borch, right? Which is another persona. >> I think we intentionally dep prioritize that for now, but I am super interested in that myself. >> Yeah. So that that's one. Uh and then there's the uh LLMs or stuff for for Borch operators, right? So being able to use LLMs to get more or make operating Borch easier >> and um we're putting together uh an approach for using uh rag. >> Yeah, I saw there was a talk about that, right? >> Uh yeah. Well, yeah, I'll be giving the talk apparently because he he because of the uh US governmental shenanigans um he won't be able to travel. So yeah, uh but anyways, but yes, he he's put it together and I'm I'm going to give it for him. >> But uh let's go again back. I think the first step should be people will be uh uh uh relaxed to put the agent on cloud foundry and connect some services and be uh uh uh sure that the secrets there won't be sent somewhere uh you don't want uh like uh sandboxing uh and we can go connecting services first we can go like when suggested with the primitives we have binding service uh maybe in the future We can think of uh specific type of service which uh for example represent MCP MCP servers or uh skills or something different. Uh uh but at the beginning we just can have a PC using the current primitives and see how it works and what does it mean >> where the friction is. >> Yeah. Um >> okay. Okay. Who's going to instruct his agent to do the work? >> So I can uh >> I don't think there's a reason why we can't all try to do that, you know. >> Yeah, just all do it. >> So Norm mentioned, for example, we now have uh because of the Proxmox CPI, we now actually all have our own labs. Um, so we can actually the only thing we don't have yet is a GPU, but [laughter] but we can we can spin up uh we can we can literally take tear things apart and you know do that. So >> yeah, I have a local Deep Seek. So [laughter] >> CPU based? >> No, no, I have a DJX Spark. >> Oh, shiny. >> Yes. I like I like the idea trying to uh figure out in parallel and then uh try to consolidate uh uh >> yeah we we can we can compare notes and then [laughter] >> yes >> you know I have homework >> yes >> oh yeah um how can we organize this work um >> I mean we we can all we don't need you just need your own repo, right? So you can just put it on your own account, create a repo, proof >> we can share our results and then figure out where we're going to go after that. >> Yeah. And maybe come back with an idea. So you basically or you just create an issue on the notes repo pointing to your for your your proof of concept that you did and put in some um and maybe we can create a proof of make a pull request and create a p uh proof of concepts um directory right so that we have some like highle thing where we capture proof of concepts that have been done. >> Yes sounds good. Yeah. Um yeah. >> Okay, >> sounds good. And uh we can open PS against this uh with the different uh ideas how uh >> yeah building stuff it's nice when you don't have to do the coding. [laughter] >> Yes, >> it's fun. Robin, I got feedback that iterations were needed on the agent generated stuff. [laughter] >> I I keep on trying to work on iterating on ways to improve the output quality of those things to be honest. >> Yeah. >> Um and it has gotten so much better in in the last year. Um but it still needs improvement. >> So yeah, I've found those balanced reviews uh the ones that you can now do with Copilot, they find a lot more. So I've for the um I'm currently working on a pull request for the UAA for um MTLS to JWT token so that you can use the instance identity search to get a JWT token. >> Um I think that's a great thing to have for agents as well, right? Like if you have that. >> So because you can use for MCP and like VIF and everything. >> Yes. Um, so if I went through maybe 20 or something iterations of of these reviews and every time it found the thing and then addressed and um I don't know it's uh find it funny that like Claude didn't find these things to begin with. But I mean yeah um um Claude has this issue that it's in what I call inly focused. If it's on the edge or just slightly outside the edge, club will never find it. >> Well, it also it's not just it doesn't find it. It'll like, hey, I see that thing, but I'm ignoring it because I'm going to my goal, >> right? >> And that's that's the real problem. So, that's that's why there's so many iteration loops and repeats and everything else. >> Yeah. And I >> open up that >> door, right? >> Yeah. Yeah. And I last week switched to using all the uh um OpenAI models again, right? Because I found those are much cheaper. So the what's it the Soul and uh Terra and Luna um those models those are I don't know interesting. >> Okay. So, but then you changed models. So, now all the things you knew about them is it's different, right? >> Right. But that's okay because >> different flaws. Yeah. >> Yeah. >> So, it's uh it's like brewing potions. You just put things together and it's like, yeah, I don't know, maybe something nice comes out this time. >> It is wizard. It is originally many cases because >> there's the only way we I I've only find AI really useful for people who already have the experience behind what it takes to build stuff already >> and getting new people to get that experience. I'm not sure how that's going to be done. >> And without that without that experience, uh what you're going to end up with is a bunch of crap. >> Yeah. Yeah. That's the one thing I've noticed is like you really need to be able to call on them like no >> right >> go look at that again you >> what you're doing basically is you're building that mental model right so you don't have to look at the code because you have that mental model of what what it's building right so you know what it's building and then you can see like but this doesn't look right and then you can say like but focus on this thing right do that different and that's yeah that's uh takes experience, I guess. >> Yeah. >> Well, right now that takes experience. I I just don't know if we can um train that experience into something. Uh because it, you know, you start with something simple, you build it up. Oh, I I saw this, I saw that. You know, it's it's kind of how humans do it, how we do it, right? >> It takes a lot of effort and it takes a lot of years of time. And the nice thing about AI is time is compressed. So we can do something like that. I mean, one of the things I'm look at uh relative to the plugins is having an AI plugin, right? That you plug something into the the CLI command and then execute do stuff and have stuff feed off of that, right? And then now you could have your your uh whatever uh tool that you're trying to build actually work through uh with the this CFCI interface and do what it needs to do. >> But isn't that more of like a skill at this point? >> It it is and it isn't. Uh because when it isn't in the case of it needs access, [snorts] right? And right now the plug-in gives you the access that you need in order to do it. Now you build that if it is a skill but that's a skill for a skill is for one topic. What you want is a skill that maybe it's going across multiple um envir um CF environments. So you you want to be able to switch between u using the CLI getting the data back right because you get the raw data basically. I I'm getting the raw data now through um CFCLI, right? Or or Cappy. So, I'm using, you know, the Cappy is the stuff hooked hooked up, but really all that we're really doing is talking to Cappy, right? In terms of getting the stuff done and how you tie that all together is all dependent >> and I saw also been seeing PC's having MCP server behind uh so in front of the CLI or they cut the APIs. Okay, >> MCP is is not what what I find is that MCP is uh quite heavy. It pollutes a lot or spend you lose a lot of tokens because of it, right? So that's why skills are usually a a good CLI with with skills is usually more efficient. >> Okay. >> So for example, I don't use the MCP server for GitHub. I just use the GitHub CLI and say like create a pull request and it does it right. So I think the same is true for CF. I actually don't have to do um I don't even need a skill. It just knows about the CF CLI. It's trained in in most of it because of the documentation I guess. Um >> yeah. Um so uh regarding some timelines uh should we try to come up with uh some puscies before CF submit and then maybe you can have a little bit time to incorporate in your talk the different ideas and then we can discuss during CF submit in uh person. Uh >> mhm. >> So have one hour or two hours sitting together and discussing those uh puscies the uh advantages disadvantages and try out to come with one uh integrated uh approach how to implement the sandboxing. Um >> I love that. That would be a really useful uh I don't know useful u way to spend our time. Right. So if we are together at CF summit, it would be good if we already have some ideas on our own, right? Because then we can at that time discuss in person and be productive hopefully. >> What I like with sandboxing sandboxing is also this can be also uh useful for normal works on CF not only for agents but uh every workload can profit from that if you can activate. So I you say okay I want to run this work quote in a sandbox mode and then um for whatever reason because it's a legacy I don't know someone gave me an application and I want to run it on CF >> I also feel like there's a great use case for when we do have the maybe even before that but when we do have the Bosch based uh LLM serving >> we can actually also probably I'm not Sure, there might be more use cases than just that one, but the isolation segments feature, I feel like it could be very useful uh for some some of this stuff too. Like if they want to isolate these AI workloads away from application workloads, for example, we can have an AI isolation segment or some some kind of logical partitioning of these things. >> Uh yeah, that there's something there. I haven't thought through it, but yeah. >> Yeah. And I like that we're not specifying sandboxing too much, right? because then we can all start with our own ideas of what that means sandboxing for CF and then we're going to probably find that we have really different ideas. But I think that's good, right? Because then >> Yeah, >> that's why I'm exactly why I'm suggesting we should all try it like >> Yeah. do it all and then like just yeah >> think through the promo of what that means for you CF and sandboxing like just really vague high level and then >> try to come up with a concrete implementation proof of concept something um >> I like that >> are we are we going to now this meeting are we going to like try to other than summit week obviously but are we going to uh try to like do this every two weeks or something or I uh uh let's discuss this during CF summit maybe. >> Okay. >> But it would be good to at least present the results of our PC before CF summit. >> Yeah. >> Uh I'm I'm I'm around until the 16th I start traveling. So before that >> we think we will be able to we have two weeks actually. Um I'm not sure. >> 24 Agents can run 24/7. You know that, right? >> Depends. >> Just need to make sure that you setting than others. >> No, no, no, no. But I can run my agents from my phone, manage them in my lab. So, >> it's fine. >> I want to uh work with and discuss and uh generate ideas on location. >> Well, we'll get as much as >> there's your there's your problem. You're interacting with humans. You should be interacting with agents. It's much more efficient. [laughter] I learned that doing in parallel and still using my head and trying to figure out things still [laughter] to do. >> Have your agent handle the humans. That's better. >> I see. I see. Um, >> so you got the split plane as well. The same problem. [laughter] >> Yeah. >> I don't know. I will write some meeting notes. I want I will just write meeting notes to people uh see what we discussed and what we want to work on and uh and so maybe others can also then in parallel do stuff I'm not sure about putting concrete date before CS submit if you if um we maybe >> I think it's a short timeline I mean it can be done but I don't think it's >> we should just present what we have or >> we can state that we if we have uh PC's ready before that we can organize on demand a meeting and then do it without committing to a date. >> We can do it async as well, right? Like if you just create the issue or the PR and then we can read it there and then maybe if needed we can have a meeting oneonone or whatever. >> Yes. Yes. >> That's fine. >> Have your agents meet. >> Yeah. >> There's this agent to agent protocol. We can just set it up. It's fine. >> But we don't have the identities yet. Uh >> the identity. Oh, we should start with the identity. [laughter] >> We just found the tension. We need identities first. >> First identity so our agents can talk among themselves and then we are fine. >> Want them to exchange secrets. I first want them not [laughter] able to exchange secrets and then identify themselves. What I what I love about this new thing though is like it's opening up the [clears throat] possibilities of what we can do by making us more a a lot more productive. So I personally spend a lot more time thinking about what the heck I want to do instead of doing it >> basically and calling [laughter] >> Yeah. >> Okay. All right. Can you put the meeting notes in also a pull request to the so we keep it all in the same repo? Just create another directory or something or something. Mhm. >> Yeah. Another folder. >> Yeah. Yeah. Yeah. I hope my agent is able to do that. >> Oh, it will be able to. >> Did Did we record the uh Did we do a transcribe on this? >> It says it's being recorded. >> Yeah. But does it does it transcribe then or can it? >> It can be. I mean, >> okay. Yeah, probably >> you got to be careful the transcription. >> Uh, whoever's the host could probably >> Oh, yeah. It's not Oh, it is. Meeting summary generate transcript is not >> It says request. >> It's not turned on. Okay, that would be good to ask Ram for the next time. >> Um, >> yeah. Okay. Yeah, sounds good. >> Great. >> Thanks. Fine. >> Now go play with your agents, folks. Have fun storming the castle. [laughter] >> Take care. Bye. See you all in Hidleberg. >> Thanks for your time. Bye. Bye. Bye. Bye. >> Bye. Bye.