Caffeine V3 is SUPERCHARGED (ICP) — Agent Teams, Skills & Sovereign Deploy (Plug and Play)

Added: 2026-05-05

[00:00:00]Yeah. So the the big shift with V3 is the agent team V2 to V3. The two main issues were the latency have a prompt and then you wait for a long time and and then it breaks and then you have the prompts that don't follow what the user is asking. So you prompt it and you say, "Hey, I want to put this over here or oh, I want to include this page on my my website." And it would not do so. So then there's two big things to fix, which is making something faster, but also making it better and and able to follow what the user is saying. And fixing one of them has to come first.

[00:00:37]Like you can't fix something that isn't following the user's directions and make it really fast. You can always improve on latency, but you know, running fast into a wall doesn't really help, right?

[00:00:47]You you need to move around it. So that's where the agent team comes in and that's where the different waves come in. Yeah. So the big thing with V3 was solving the the agent being able to understand the user and also help the user get out if the build is stuck or if the build runs into these issues. So looking a little bit into if we have the flow of the linear pipeline we have single agent that goes through and you have the plan, you have the build the back end, you have the build the front end and then a quality check. you have one agent that has to do every single step. And one thing about agents is context is king. So what is context?

[00:01:30]It's like your the code of your project, the prompt that you're doing, images that you put, you're kind of all of that is context that you're feeding into the agent and that goes through this this whole pipeline. So then with the pipeline approach, if something happens where there's a mistake, so let's say it's on the the back end. So building backend, let's say you you got your application and you're really happy with it and then you do an update and you say, "Hey, I want to add this to to my application." And it goes through the pipeline again and the the back end isn't able to produce it. let's say it was a bit more complex or or something happened in that build process, the entire project would fail in in V2. So, and it would be very hard to get them unstuck. I think a lot of users were >> brick walls.

[00:02:25]>> Yeah, exactly. And it was one of the the major complaints was, you know, I've put in so much time and effort, right? We have people that are, you know, each deploy that you have is a version and and you can go back to it. You can go to the next one and and people were on version 400 just, you know, incredibly complex apps that they're building with caffeine. But again, in pipeline, you you have kind of a to be one shot through. If anything happens inside that process, you have to go through again and and loop back and and and try like you can't stop here in the back end and say, "Oh, let me just work on front end." And then you would have V2 that says, "Hey, it's it's done. Here you go." Which isn't ideal. And when you have the the pipeline approach, it's less of an agentic one. And it's more of a, you know, let's go through A, then B, then C, then D, and if something breaks, we have to go through A, B, C, D, and E all over again. So V3 kind of opens that up to an agent team. The Kepler live stream was really good of breaking through exactly, you know, what does the composer do? What does the the PM, the discovery agents, the design, backend, front end, QA review, and QA visual. So all of these and it's it's more of a at this point in time that the user is building this project what do we need to do which was a question that isn't possible on the pipeline aspect of like I guess agentic building but here we can go through one agent that then talks to another one then talks to another one talks to another one and you can you know loop through those you can have a pipeline approach within the the multi- aent. So, you know, you don't want your front end to start by uploading things before it's actually done the skeleton of the project. You still do want pipelines, but at the right points.

[00:04:28]>> So, it's a bit more modular in its approach.

[00:04:30]>> Yeah. Yeah. A bit more modular and and and working well in terms of what people have been able to build. A lot of people are saying they can build more complex applications. A lot of people who were stuck on on B2 >> have gone back and go through and there's still some, you know, pain points and and scaling. I mean, we >> launched and then a lot of people kind of flooded through and and had to do some some nice load balancing on there, but we're eager.

[00:04:58]>> Yeah. Yeah.

[00:04:59]>> Eager to build. But it's >> just really quick, Swissy, one point that raises that comes to mind for me is so many of these people who did get brick wall building very complex applications. So now they're able to revisit those apps with V3 and >> yeah, >> ideally re revive them.

[00:05:16]>> Yeah, I think there's some that are like if if they were kind of day one or or very very you know old projects might have a bit more issues going in. Okay.

[00:05:27]>> But there is a path that that came with this update that other users, you know, that they shouldn't have any of those issues. Yeah. And then you can you can go through and >> kind of prompt and push its its limits.

[00:05:39]But again, now you have in that one prompt the entire agent team, but one prompt now kicks off a team of agents to work on your app.

[00:05:48]>> I'd love to hear a bit more about this multi- aent flow.

[00:05:51]>> Okay.

[00:05:52]>> And maybe you can define for the audience what exactly that looks like.

[00:05:56]Is there a set team of agents that works on every single app that's prompted in caffeine or are certain agents called upon based on the specific requirements of a particular app that user is trying to build >> a normal like workflow let's say if it's B2 normal workflow >> it's blackboard time >> blackboard time >> would be you have your your prompt and then you have your output that you're hoping for and then your workflow is is straight straight straight through. If something bad happens in here, you're you're screwed. You have to start again, go through and and output like that. And maybe you got it right on the second try and you're happy. But then B3 brings this approach of you have your prompt and you have your your output. So that doesn't change, right? Like you still when you're app building or you're you're prompting and all these things, the big difference between like cloud code or more of this agentic engineering is more step-by-step building. So you might say, "Hey, build me a web app."

[00:06:59]And then it builds, you know, let's say it goes on Versell and it it goes on whatever is maybe trending or in its like pre-trained data that it's it's popular, right? It'll probably do Nex.js frameworks and all these things and and trying to stay a bit high level here.

[00:07:16]But you would have to then, you know, build that up and and you might get something that's just local. You might get something that doesn't even work.

[00:07:24]you have to iterate through and you can iterate quickly but but you still have to go step by step by step whereas on caffeine you have more of this app builder approach where you can prompt and then you walk away and that that thing that you've been working on is is built. Um so in V3 what we have is these kind of multi- agents that are able to talk to each other communicate and also interact and break certain things down.

[00:07:52]So you know let's say this one is the the background agent backend agent sorry and then front end agent and you can this opens the door to like a parallel parallelized approach. So when you have let's say an agent that's working on back end you can work on some things on the front end. V2 had it where it was only one after the other. So sometimes people would add something on the front end, you know, what your your site looks like and then all of a sudden it would break because you added maybe a you tried to add an admin panel, right?

[00:08:27]And you wanted that visual, but it it actually touched the back end and then it would it would break and you wouldn't be able to get out of it. So let's say it goes through and then it can talk to the design one and then it can talk back to, you know, an orchestrator and then you get your your output. what you save is the the context that we started with which is you know you have to keep in mind all of these things that the app is doing. So how do you know what the design looks like? How do you know when it's done? How do you know when when all these things happen is you have to take that that multi- aent approach because then you can have the the the context managed a lot better. The other advantage that you get I guess is different models different things that that people produce. Let's say if like Gemini or whatever is is doing good on on UI stuff, then we can get that in there. If we want certain skills at certain points, we can have agents that have those certain skills. So, you can have >> nice >> like your your experts that have different things and they all come together and and give you your nice looking UI, you know.

[00:09:38]So if you wanted to poured in some UI design expertise from Gemini, you said that's possible.

[00:09:44]>> So you would I mean it's possible in in terms of the caffeine kind of builder process, but I think skills that other people have done I think will will look to to to add that in. The next big big thing is is going from V2 to V3 was skills in general. Be before it wasn't even possible to to have skills. So, so the shift is is definitely monumental in that sense. You have V3 skills that we've added and now you know it's kind of curated through caffeine. So, so we did kind of the the work of seeing what are the best skills out there, what are their outputs, how does that relate to ICP and how can we curate kind of the the best of the best for for people to to just come into caffeine, right? Like if you have a side by side and you have clawed code and you have caffeine let's say V3 um here you have to have the user that goes in and finds the the skills >> and finds the the the different resources >> or prompt cloud code to find the optimal relevant skills.

[00:10:52]>> Yeah. It takes that manually.

[00:10:54]>> Exactly. And then like >> a big one is like ICP will say cycle management.

[00:11:00]>> Mhm.

[00:11:02]>> Right. And if you if you've never interacted with with this part, you know, it's it's kind of where things start to break down a little bit. Um and then V3 is the user comes in and then prompts.

[00:11:16]>> Skills are baked in by default.

[00:11:18]>> Yeah. Yeah.

[00:11:18]>> Massive.

[00:11:19]>> Which is which is huge as well. V3 as well. status bar. Yeah, there's so many there's I think Christopher had posted where it was here's what's new with V3 and it's >> Oh, yeah. It's in a ton.

[00:11:30]>> Ran out of room.

[00:11:32]>> Ran out of ink.

[00:11:32]>> Yeah.

[00:11:33]>> Yep.

[00:11:33]>> Which is true.

[00:11:34]>> Yes. And just what are what highlights out of those upgrades and new releases come to mind for you as being most impactful?

[00:11:41]>> Yeah. I think well getting unstuck is huge. like having an an agentic system that is able to follow what the user asks I think is a if you don't get that right then then we can't crucial yeah we we can't upgrade we can't move quicker there's no point in speeding things up if if what we build isn't isn't good and then skills is also huge because we can curate kind of the the industryleading design sense into somebody who's never done design never done frontend and be able to have that kind of curated knowledge into their application building. So skills is big, getting unstuck, having the agents is big, agentic versus pipeline. So again, we we kind of brief went over pretty quickly which was the the pipeline where you go from A, B, C, and D. And if something happens, you have to go all the way through A, B, C, D, and E again.

[00:12:35]>> Yes.

[00:12:35]>> But agentic is kind of like, okay, A to then where should we go next? And then, okay, we did this, but now we should probably change the back end again. Or, oh, we changed that, now we should change the front end again. And then, okay, now let's check all of what we've done, and then we'll say we're done. So, you can you can bounce around that a little bit. And >> so, the order of operations is adapted.

[00:12:59]>> Exactly. Yeah.

[00:13:01]>> Based on the need, maybe it's good to go through all the types of agents within caffeine. the the blog is a really nice source for anybody who's looking for more information.

[00:13:12]The so we have caffeine, we have composer which is kind of the orchestrator. We have wave one which goes through discovery PM and gives contracts and learnings. Right? So that's the contracts and learnings aspect is the one where we had the web design one that was going back to the composer. That handoff here is where you you're not sending the full code. you're you're sending just what you need to get from wave one to wave two and you're saying hey did you do your work like tell me about it are you good to go to the next step yeah I'm great or you know hey did you do your work and it's like hey I'm the front-end agent I did this work but you know the backend agent might not know what what they're supposed to be doing and that's where you have you know the the typical I guess PM that that comes into the the fold as well where you can have your your PM am that goes in between them and is able to see and is like alert alert you know we we we're missing this backend aspect and before in the pipeline you didn't have this at all >> you didn't have that that kind of >> context manager >> yeah [snorts] yeah >> exactly and then there's some nice things as well like quality review these things down here let's see yeah the might be a bit small but review and and visual is is where you have another check before something gets output. I think there were a lot of times in V2 where we would have the generation happen and it would say, hey, I'm all done. I'm good to go. But now we have a check that goes through and sees, okay, did this actually did this actually happen? Did the users prompt get kind of implemented properly? But that's for for V3. big thing is just this foundation of multi- aent being able to add skills to them and being able to improve them versus more of a a hard-coded pipeline workflow approach which you can perfect right if we we go through the trajectory of where v2 would bring us you know endto-end builds that are very very you know scoped and maybe we we nail those perfectly but as soon as people add complexity we wouldn't be able to support that so the journey from v2 to v3 Three was completely taking this this foundation with what we've learned and and all the interaction and feedback that we've gotten from people is is substituting that in for something that can now scale and and be this this bedrock for agentic foundation of you know do we want a marketing agent for caffeine users? Do we want an email agent? kind of all these different things that that we can now unlock because of that that foundation being made, changing models with with different updates, but also making things quicker. I I think a lot of people might realize like, hey, you know, I'm very happy with what I can do with B3. I can make the most complex application that I that I've dreamed of, but I want it to be faster. Like it's not fast enough. I I prompt and then I'm waiting, which is true. I I think trying to solve all of those uh as we've done might have been a bit too much of of that that hurdle. But again, the the pace at which caffeine is developing is slowly getting to that flywheel point which which I can do another breakdown of. But with V3, we're we're able to now go through and say, okay, where can we speed up the the PM that talks to the the design agent? Can we spawn right the the multi multitude or whatever of design agents that are able to each take one thing else or front-end agents that are able to do things in parallel. Can we also take the the context that we have and rather than giving the the whole codebase wi-i which we still have to make sure that there isn't those issues that are happening can we go into here and say hey all I need is this chunk and then give that to the agents and then they they're able to to continue producing good outputs. Essentially, we now have the opportunity to go through and chip away at that latency, chip away at at the things that that we can improve. There's I mean, you can go into speculative decoding and and a bunch of different abilities. You can also improve skills. You can see how people are are using things and say, "Okay, this skill needs to be improved," or, "Oh, we're we're losing a lot of time when this model is doing generations.

[00:17:48]So, is a different model better? Is is is there context that we can drop?" Kind of goes through through all these things. Phenomenal plugandplay deployment straight to the blockchain.