BUILDING WITH ANTIGRAVITY 2.0: THE UPDATES YOU NEED TO KNOW

Dalton Anderson (00:01.314) Welcome to Venture Step Podcast where we discuss entrepreneurship, industry trends, and the occasional book review. I'm your host, Dalton Anderson. In this episode, we're going to be doing a deep-dive into Antigravity 2.0 and Antigravity Command Line. These were recently announced products from Google's Google Cloud Next that I talked about a couple weeks ago in a podcast episode. If you would like a full deep-dive breakdown of what was announced and

why is it important, please listen to that episode. In this episode, we're going to be discussing these agentic coding tools and how to get the most out of them and why I'm so fond of them. So, first off, before we dive into the tool and share my screen and get into a live demo of some stuff that has been built, I first want to backtrack and discuss Antigravity 1.0 and 2.0, and then Gemini Command Line and then Antigravity Command Line. First, let's start with Antigravity.

So Antigravity 1.0 is a fork, which is basically a copy of this popular coding IDE called VS Code. VS Code was an open-source or is an open-source coding editor, IDE, whatever you want to call it, that was originally created by Microsoft and has been supported by Microsoft forever. Like when I was in college, I used it.

And it's a phenomenal tool. It's got the the best resources. It's Everybody knows how to use it. It's standard.

It's open-source as well. And so there has been some popular takes where people have been taking this coding editor and taking that and flipping it to make Cursor, which is another agentic coding platform. And then there is Antigravity, which was a fork of VS Code. So same thing, a copy of the VS Code to create this new product.

Dalton Anderson (02:08.118) And the innovation of Antigravity was really this thing called Mission Control. And Mission Control was an extension of this IDE. And an extension in this instance is a new window that opens up when you say open Mission Control. And Mission Control allowed you to manage multiple instances of agents. And these agents could create subagents and one workflow that you had

ongoing could have twenty or so agents running at the same time and you could have up to fifteen workflows or instances open at the same time as long as you had enough local or cloud compute and enough tokens to support that. That's either your usage or your wallet. Either or because it can get expensive if you're not optimized correctly.

So Antigravity 1.0 is a extension, whereas Antigravity 2.0 is a platform for agent orchestration. Where before it was a new window, now it doesn't even have an IDE part of its instance. It's a separate thing. It's a separate product. You have to download a separate Antigravity IDE, whereas Antigravity 2.0,

it's just the orchestration part. It does everything on its own. You don't see it. It can provide you demos and screenshots and videos, but you're not actually in the codebase if you don't want to be, which is completely different than how things would work typically to get the most out of these tools. There is terminal-based coding, which has been around for, I don't know, a couple of years now.

There is Claude Code, there's Codex, and then there was Gemini Command Line until it was retired. And now it's Antigravity Command Line. The main difference between Antigravity Command Line and Antigravity, I think, is the speed and architecture of the tool. So Antigravity Command Line is built in Golang. Golang is world-class at information orchestration. So sending

Dalton Anderson (04:31.603) a package of data from one place to another place and receiving it back and sending it however the pipeline is, it's the best. This whole reason Google built it was to be able to send data from one place to another place. And that's all it's really good at. I mean it's good at other things, but that's the main thing. It's world-class at data orchestration and sending things back and forth and receiving information. It's by far the best.

And that's what Antigravity Command Line is built in, which is great. And then there's some other stuff that's really cool about Antigravity Command Line, is it has a full-on toolkit built into it. I mean, Gemini Command Line had some of this stuff, but it wasn't as straightforward as Antigravity. I know people like Gemini Command Line, but people just don't like change. And there are some features that aren't within

Antigravity Command Line that are built out in the 1.0 release, but in a general sense, the process is faster, it's better, the UI is better, the functionality is more smooth, it's more straightforward, it's more user-friendly. Overall, it's a better product. So I think it's a win-win for everybody. So those are the two tools. So why don't I just share my screen and show you what

Antigravity 2.0 looks like. And then we could talk about something that I decided to build and walk through some of the features. Because the features set is kind of similar within the Antigravity platform, I call it, and then Antigravity Command Line. So share my screen. Let's see here.

Dalton Anderson (06:38.239) Alright, so you should be able to see this.

Okay, so within here, y you have all your projects. And your projects are structured wherever you want them to be, but typically people put them in their home directory, which is depending on where your system lives. Like if you're in Linux, Mac or Windows, it's in your your user. So it's in like your root of your user, which is referred to as your home directory. And so it would be whatever your username is, a folder, it's

lowercase dev and then you put your project in there. The historical context of that is back in the day, there the way that you ran code was that you'd have the source and then the file. And then it was the the like this user slash source in within your computer, because not everyone had home computers, so this really wasn't a thing until Linux came out. And then Linux had

users and then within there you would sandbox your system and so you needed to switch to this user split out because everything was sandbox so that's why it moved to the home directory. It's very interesting. And then the dev folder thing it's just like a natural progression of development. And so it went from source, it went to dev.

And that's that's kind of where it lives. And then all of your coding stuff and your agentix stuff lives in your home directory. And that's just how it is because it's easier to sandbox and provide a

Dalton Anderson (08:24.884) ease of permissions within that folder. And so it's just much easier. Okay. So this is the UI in here. So I was working on something that I was inspired from me trying to build content for the podcast, Venture Step. And I was thinking about how can I create something that's useful to myself, which isn't wasteful, but also interesting to you.

And I was like thinking about all these things. I last time I did some kind of claims thing related to insurance, which maybe is a little abstract for a wide audience, but people can get behind building a system for the podcast. You can reflectively look at it and think about your hobby or things that you're trying to do. And it's like, "Wow, like if you could just do that by asking questions, like how can I do my thing? Like I've always been interested in X.

How can I take the content that Dalton created here and follow this framework?" And this episode, I will be providing the same framework. I just I can't help you do the questions. You've got to do the "what," I can give you the "how," and I can give you the "do" kind of thing, or I can give you the framework and the execution. You just have to download it from the internet

and install on your computer, which I'll I give you the commands, it's on GitHub, and I'll link it in the show notes. But that's the thought process. And the way that I was inspired from this, and maybe if I backtrack a little bit, the way that I was inspired to build this platform was I was building out this LinkedIn carousel. And on LinkedIn, the go-to to really push for engagement

is these LinkedIn carousels. People really like that because when you put a PDF on LinkedIn, it splits out the pages. And so you basically create create a slide deck and you have people swipe through that slide deck. But since you're swiping on it, there's a lot of engagement rating from this carousel, and people find it interesting. And a lot of times if you're making the time to make a carousel

Dalton Anderson (10:48.044) and put it out there and post it and name the file, it's normally high-value content, which is where it thrives on LinkedIn, because LinkedIn people want to feel like they've got some edge or they're they're they're on the forefront and this carousel might make them feel that way. So let me share my screen of what I was able to build in Gemini. Okay.

So this is Gemini Canvas. And so Canvas can build apps, it can build word docs, it can make documentation, it could do a whole bunch of stuff. And I think it's a really good tool just to prototype something to really think about where where you're at and what you want to do with your idea and just prototype before you do anything serious. And so I was trying to create a

carousel and I was thinking about the prompts and the thought process and I threw it into a couple things and I was like, "Okay," and then I was like, "All right, let's try to just make it like who knows." And then it just made this and so I thought it was pretty good. There's not as much editing as I'd like but this is a really good start to where you can add slides, you can change the template of

your slide so you can make it a quote, you can split it out for contrast, you can make it a primary hook, you can change the header and title, but you just can't change the body. You can change the colors, but there's not as much detail as I'd like. But overall, this is really nice because then you can pretend on what like what it would look like on LinkedIn and

it's just great. And you can flip through. So this is what it would look like to a user when they're on the platform. So I thought that was awesome. I was like, "Wow, what if I could do this but on a higher scale?" And one thing I have an issue with is some of the things that I'm trying to do with Venture Step require too many tokens. So on this site or on a website like chat.com for ChatGPT

Dalton Anderson (13:10.742) or Anthropic or Gemini, you only get a a certain amount of tokens on the website. If you want a lot of tokens, you need to go onto the API. And the API costs money. It's not part of your your normal subscription, et cetera, depending on your subscription. Like Google has it set up to say that you get a hundred dollars of Google Cloud credit.

And your Google Cloud credit is also part of the API. So you do technically get $100 in credits. I think they got rid of that. I'm not sure now that I'm saying it, because I know that they lowered the cost of the subscription subscription by like $150. So I don't know if they're still providing that large of a credit. Regardless, what I'm saying is if you want to do a lot of token work, whether it requires a lot of context,

then you have to use the API. So it's like, "Okay, well, that's awesome." And then I'm I'm getting help. Like Venture Step is not just me anymore. And it's easy for me to manage this content, but now I'm trying to make LinkedIn carousels, I'm trying to make thumbnails, I'm trying to upgrade the content. And so for that to work, you need to have some kind of prompts. And then you need to have

that person who's helping you, they need to have access to AI. And that costs money. And so you either have the option to make an account and give them a subscription, or you can create a platform that has this API integrated and they can use the platform and then call on the API, which I think is easiest because I I needed it anyways, because some of the work I wanted to do wasn't available because I needed more tokens. And

Just to write an example, like a blog. A blog is two thousand to eight thousand words. And then from there it takes a subsc the transcription. What I would really want is a blog. And then underneath the blog, it's a transcription of the podcast episode. What that will do will that will provide a massive, massive SEO slash geo push where it's just a massive amount of content.

Dalton Anderson (15:36.575) Users don't want to read the transcription. The machines do. Users want to read the blog. But if you can put them all into one, then it makes a lot of sense. But I personally don't want to format the transcription. AI can do that just fine. And I personally think that if it could take the transcription, format it, and provide some context, it can create a good good enough blog, and then it can be edited by humans,

where it gets it to ninety percent, eighty eighty five percent or something, and the human could spend an hour getting the blog in a good state. All that being said, I worked on a platform. So if we go back to this tool, stop sharing, show something else.

Dalton Anderson (16:32.398) All right, so if we go back. So this is the platform. There's a couple of interesting things. So you will be given these things called artifacts. So there's implementation plan. So it could tell you what it's trying to do. There is tasks. There is how it tested its authentication. And then there is a walkthrough of what it did. And then it tells it what it's what it's done. So I was trying to create an authentication. So it was testing it and it was failing. And I was like, "What's going on?"

But regardless, so this is this is the overview of kind of what it's doing and I have all the tasks and you can see all the artifact and and you can go through and look at them if you want and then it can tell you what's going on and what's been changing and if there's any sub agents running at this at the time.

So it has native it is you can't see or you can't it has native audio audio recording, which is great. And then you can have commands. So there's actions and then within here, there is actions. So you are able to set different things according to what you want to do. So a goal is what it sounds like, a goal.

I want to I want to run five miles. In this instance, if you're saying that to the agent, whatever you're trying to do, it just does it until it can't do it anymore. And it finishes the job. So it will run until it's completed. It could take a couple hours. It It could take some

some X of time, it could not take very long, but a goal is something that runs agentically where it doesn't necessarily ask you for little things. It just makes the best decision that it thinks that it can do. And it's important to have the right context and the right setup when you're utilizing this, especially if you're trying to build something out and you don't want to be disappointed with it having taken a different direction.

Dalton Anderson (18:47.724) And then the next thing is you could schedule tasks, and then you can schedule tasks both on your local computer and on your cloud, depending on how you have it set up. And then the browser is what it sounds like. It It does the browser. And then /grill-me is one of my favorites. So /grill-me is something that you can have activated. And we'll see if this does voice type while I'm recording. Hello, testing, testing. Okay, so it works.

Okay, so we have this platform and we recently just finished the authentication. What are some things that we need to change? So when we think about the user flow and how they interact with the content, how are they going to specify which thumbnail is good or how will they be able to provide feedback about different things that are being said and how will they be presented different options when the content is generated?

Dalton Anderson (19:51.693) Okay, so that just went so it's on /grill-me. So /grill-me is a function which literally grills you. And so it'll ask you a whole bunch of questions. It's an interview. It'll It could be multiple choice, it could just be fill in. It It kind of depends on what it needs. But in this instance, we'll see what it says. Okay, so it it gives you a notification as well. So it says, "How would you like the user flow for selecting and refining

thumbnail graphics?" And I'll just read one of them. So recommended. Implement a three-variant concept grid where you view three generated visual ideas, render background views for them, and edit one to c to edit. One one click to edit, refine, and export, if I could read correctly. That seems fine. But

I think it'll be a mix between I think it is a mix bech f with let's say from one and two two when

User has selected and finalized

Dalton Anderson (21:19.563) thumbnail, I guess.

Dalton Anderson (21:28.79) Alright. So see what it says. So it might ask another question. But it will run through all these questions and

Yep, that's fine. Makes sense. and this is how how should it handle adjustments and alternate options when generating text content? Show more notes, threads, captions. And the main recommendation was just make it shorter or make it longer or include this. So just like tactical edits.

Dalton Anderson (22:09.26) And this was saying, "How should we handle violations when we don't follow our prompts?" And so it's just gonna do its thing, it's gonna be working, and I'm gonna wait for it to finish and then

Dalton Anderson (22:29.3) I will show you the platform.

Dalton Anderson (24:48.505) Wow. This is this is taking a while.

Dalton Anderson (26:02.819) This is a hoot.

For sure. Wow.

Dalton Anderson (26:44.834) It's battling.

Dalton Anderson (27:35.073) Wow, it's still going.

Dalton Anderson (27:45.849) really didn't anticipate this many changes from what I was saying, but I guess it needs a whole rework.

Because I personally thought it'd be something simple, but I get how you'd have to change this stuff. And I may have rambled a little little bit. I really just wanted to kick something off which would be meaningful, but I this is too much.

Because we're live. We're live on the show right now. And I don't need to stand by for twenty minutes why this goes.

Dalton Anderson (28:37.751) Mm.

Dalton Anderson (28:48.941) Well, it's almost done. I think, hopefully. So the implementation plan, it describes what it wants to do with what you said. So it's taking all your information and say, "Okay, here's what I'm thinking." And it's similar to Antigravity 1.0. You can put comments and it will change the approach and it will change it live. So even if it's already running, it's already done that, you can make a comment and it will submit it and it'll make the change.

One thing that's a little different with Antigravity 2.0 is you can manage your permissions by project. And so I have it in turbo mode, which means it just does what it does. It disables all safety barriers for maximum iteration velocity.

So I'm good with it. I've always been a YOLO guy. What can I say?

But this is their version of YOLO. And they feel like this was new because prior it was something else. And then I was saying, "Hey, I wanted to do YOLO." And I set up all the settings and it just wasn't doing it on the first couple days it was out. And I was like, "Wow, this sucks," because I've got to approve everything, which is a lot when you've got multiple things running at the same time. Like if you had three different

projects going at same time and you've got to approve everything. You are just sitting there like nonstop, just do, doom, doom, doom, dun, and you can't get up. Like you just you're just a clicking monkey. And I don't like that. And it doesn't seem very efficient because I I want this to kick off. There's a question.

Dalton Anderson (30:41.335) Sure, why not? Okay. So

Dalton Anderson (30:50.285) Yeah, that's fine. I need to edit this.

Dalton Anderson (30:56.375) Okay, how should users be presented with a select between different options during the initial content generation?

Yes, toggle before and voice tones.

Yeah, let's just do this one. It makes sense. No, I could change it if I don't like it. Later on.

But that makes sense.

But man, this thing has been working so long.

Dalton Anderson (31:29.231) What have I gotten myself into? But anyways, let's go back to this. So one of the things that's special about Antigravity 2.0 is that you can set up your permissions by project. So this project could be different than another project. It has its own customizations, and then you can

just specify the different importance of what you think. If you were wanted to have access to certain things, you can remove it. But then for a different project, it could be contributed completely differently. So that's cool. And then you can set up your permissions and different things, and it's similar to Antigravity 1.0 where you can edit the MCPs available. All that stuff is pretty similar. And then you can create custom stuff on your own.

Which we'll talk about later on in the episode. If this thing ever stops running, and we might have to flip over and flip backwards and just go terminal. Let's go terminal, stop sharing.

Share something else. Let's do this. Let's go window.

And let's go entire screen. Let's do this. Boom. All right, so while this is running, we're gonna walk through terminal.

Dalton Anderson (32:52.919) So within the terminal you have access to Antigravity Command Line. You've got to download it though. So there is agy and I'll pull up. So this is gonna be Antigravity Command Line. I have an alias set up for it to be automatically set up to be YOLO. So I have agy hyphen you only live once.

And this is going to be the same thing as agy, but this is going to be agy with YOLO, because I'm a YOLO guy. And you could set it up however you want. I figure it'd be better to make it very clear that I am in YOLO mode at all times. So let's set it up to look like it's in YOLO mode. So YOLO has a couple things that you need to be aware of.

YOLO has full rights. It just does whatever it wants to do. That can be dangerous in certain scenarios. So you have to be careful on what you're doing and what it has access to. And so I give it YOLO permission, full access to the dev folders. And the the within the dev folders is full access. And it's sandboxed. Otherwise, it doesn't have full reign to do whatever it wants. So that's important.

And then another thing that's important about YOLO is you can get caught up into these agentic loops. And if you're in YOLO, it doesn't necessarily know when to stop. So you need to set up maximum iterations. And I have it set to twenty-five. But if it gets stuck, you need somebody to kill, like have a kill switch. And if

you aren't there to kill the switch, you've got to have some kind of automated kill switch, which that's what the max iterations does. But that's YOLO. But this is j almost a Gemini Command Line. Antigravity Command Line, you can access things with the forward slash. And so there's a couple cool things that we'll go through. You can see the agents that you have corrected or have on. If you have agents made, you won't be able to see your agents

Dalton Anderson (35:12.612) within here, unless you have something going. And not all instances will they activate your agents. And if they're sub agents, you can't see them at all.

So you'll only be able to see the director slash like coordinator agents. And it kind of depends on what you're doing. The way that I have it set up, it has three like agents. And we'll go over that. But you can't see them unless they're running. So just keep that in mind. And then there is the artifact. If you have something ongoing, you can view the artifacts similar to how you can view the artifacts within here. And then, the by the way, it's

quite cool and then the context is nice and I also like the quota.

Dalton Anderson (36:05.751) You You can type stuff, but there's quota and you can see how there's /grill-me, there's schedule, it's all that stuff. But the quota or usage, you can see how many tokens you have available. So I recently ran out of tokens earlier today, like thirty minutes before we had this podcast episode. And so you can see how my tokens are slowly f filling up, like a little gas bar. But I was bummed out when I ran out of tokens with 3.5 Flash because it's such a good model.

But you can see your usage and your tokens. And then if you run out of tokens, you just type model. It's kind of hard to type when I'm like presenting and doing this. And I have to keep my face completely straight, try to look at the camera. But you can just switch to a new model just fine, no problem. Right now I'm using Claude Sonnet 4.6 thinking. And you can just toggle in between models. And if you want to look at what

that model is doing when you're running something, you can just use context as a command and then it will show you what's going on. So right now it says out of "zero out of two hundred and fifty thousand tokens" and this and that. So let's do hopefully it doesn't crash my computer.

Did I lose it?

No, I I didn't lose it. Cool. So let me just take this.

Dalton Anderson (37:33.943) Dalton Anderson (37:38.319) And then we'll throw it over.

Dalton Anderson (37:43.119) All right.

Dalton Anderson (37:47.309) Okay, so now that we've got this extra stuff here, we're going to

This over here for now and I will copy and paste this. Boom. And then it's just gonna run.

Dalton Anderson (38:08.523) And so with this running hmm.

Dalton Anderson (38:26.297) So with this going, you can ask a command, say, by the way, "forward slash by the way, what are you doing?" And then it'll ask you. And then you could say, "context." And then you could see what is taking up the context window. So you could see right now it is filling up. It's right now it's thinking it has 41,000 tokens out of the 250,000. It is starting to kick off some of these processes. It's got user messages. User messages are not user messages from you.

User messages are the... it's weird the way that they have it, but like user messages and agent agent responses are part of their correspondence with other agents, if that makes sense.

Dalton Anderson (39:18.201) So it just goes and I like this little view. And then you can ask it what it's doing and it will tell you. And from there it will just run. And so this is just a goal that will run until it's finished. I gave it some pretty big projects. It's probably gonna run out of tokens, unfortunately, but it is what it is, you know. But Gemini Command Line is I said Gemini Command Line. This isn't Gemini, this is anti gravity. Come on, man.

But this is a really cool tool and I like it. It's very user-friendly. It It's super easy to know what's going on and how to use it and to change and it makes a lot of sense. It's I think it's great. I think that this is a really good first release. I mean, it's an iteration on a previous product. but interrupt.

Dalton Anderson (40:28.141) Okay, so with that running, this is done. So let me go back and let's share the website. So the website, remember, it was inspired by this. So if we look at this website, and if I reload this, maybe it will won't log me out. Okay, cool. So you can come in here and you can put in an episode. And so, we of course, I'll select the episode that's related to this episode, which is "Google Cloud Next: Scaling Autonomous Workflows."

And unfortunately for myself, it doesn't look like I can copy and paste.

Dalton Anderson (41:08.815) I guess we will just

do this. Apple Podcast. I know Apple can allow me to cop copy and paste. All right. So then we'll take this. And so I'm just gonna fill this information out quickly. Boom. And then I'll put in the Apple link, of course. I'll put in the Spotify link.

Dalton Anderson (41:43.871) And then I will take the YouTube link, I guess. Let's skip it for now. And then I'll take the transcription, which I do not have. All right, let me get that.

So once I get the transcription, I put this in here, it's going to do some cool stuff, which I want to share. And this took me about an hour and a half. Nothing serious, just vibes. And it's really cool. It probably would take you longer if you were inexperienced, but if you're super experienced with this kind of stuff, then it wouldn't take very long at all. But it's definitely just a viewpoint of like what's possible.

In a short amount of time and just giving you inspiration of, "Hey, I've got this side project or this thing that I've liked or I wanted to do." And here's the opportunity to do something. And it doesn't necessarily need to make money. It could make your passion easier, right? To do, which is what I'm doing. And so if you download this, this is a transcription and it is only twenty eight thousand characters.

The ones with guests could be like 70,000. So I create the episode. Episode's been created. I'm missing one of the links, but it's fine. So then we'll go back to the studio desk and then we will change this episode. And now it will populate. So I click show notes. And then the show notes will populate within here.

Dalton Anderson (43:28.836) Boom. And then it's gonna pop up on the YouTube description. Boom. And so these platforms have different takes. So YouTube is more of a search engine than it is a social media platform. Like people interact with the content, of course, but how do you find the content? That's really through searches. And Google aggressively uses the YouTube content on Chrome and other search engines.

Whereas Spotify and these other types of platforms, majority of people listen to podcasts on their phone. And so you need something that's phone-friendly. And it for some reason this has markdown, which is not supposed to. So that's interesting. That's a That's a violation of the prompt. It wasn't doing that yesterday, so I'm not sure what happened. But this is done via transcription.

And then these prompts that are in the background within the prompt manager. And then it populates everything. So if you go to social media posts, it can do the same thing. And the first thing you gotta do is extract like the main takeaways. And so the main takeaways would be,

you know, they're launching new products, Antigravity. Let's see what it says. So it says, "a multi-layered approach to agents." Yes, that's true. Key operational is it's its own first customer. Yes. It demonstrated the live demos. I I did say that. The company aims to lead the race in AI by focusing on superior context, availability, and compute infrastructure, leveraging its extensive data from Chrome, YouTube, alongside new specialized TPU chips. Yep, I said that as well.

A practical application was shown with Google Workspace Studio, automating podcast guests, emails, workflows, and summarizing content. Yep, I did do that. So all this stuff is legit, right? And so the what I'm trying to do is I'm trying to have a full context window when I'm creating social media posts. So you have the transcription, then you have these key takeaways. So your transcription is the finite grounding where it's like this is the foundation.

Dalton Anderson (45:45.963) And then these takeaways are the floor. And if you're creating social media posts or you're doing stuff, you need to be grounded in each. So you shouldn't have more stories than you do floors. And so this instance in this analogy, we've got one, two, three, four, five, six floors. So your building can't have ten stories. If that makes sense.

And then this section you can toggle between who your voice is, but we're gonna do Venture Step, so you can do like this. And then LinkedIn is really about takeaways and firm,

Dalton Anderson (46:32.016) "Wow, this is this is really good stuff," where they feel like they're on the cutting edge. And I feel like LinkedIn is more of a like buckshot where it's it's got a small spread to it, but it's got spread. And then

Okay, did a good job with this. Wish it put more information, but it was providing more last time. It's weird. But with X, I would describe X as a laser. Like X is as little amount of words and the most interesting takes in the sh shortest context window you can provide. And so this is what it came up with. YouTube community.

It It does everything within this one click, it batches it, which is nice. And then this is the studio that will create

the threads. So it automatically creates the threads for you, which is great. And so it it it creates a a thread to put on X. And then you can specify for Dalton or the brand. Right now it's the brand. And then you can do carousel studio. And so you can create a carousel similar to the thing I showed you before. And you could just click, you know, regen the storyboard and then it'll create

The LinkedIn Carousel Studio, which is more expensive, so I'm not doing it. And then this is the clips section to where if you have a clip, you can put in the transcription of that clip and it'll help create a caption for it slash description. And this is how you could create thumbnails. And then this is how I would create blogs. So this is an easier way of getting a first pass of what

Dalton Anderson (48:29.156) the caption should be or description should be versus doing all this stuff from scratch. Like it just not, it's just really not feasible for someone to make all this this content easily without a big struggle and do it consistently without some kind of structure. So this is just a private structure. It's a work in progress. It's something that I'm trying to build out

and provide a structure for myself and the team to where when we're making stuff, the stuff is consistent, same brand voice, it is somewhat automated. And then the users, or not the users, but the user of the platform creates the right voice per se. Okay. So that was Venture Step footprint. Okay. So this is what I built Friday night

in like an hour and a half. The next thing I want to talk about is getting the most out of these tools. So I created a GitHub repository for people that are interested in learning more. It's pretty structured. I helped AI remember all the things that we did together and updated and then put it all together within this repository so other people could use it.

So the main thing that this does is it it restructures the way that it calls agents and sub agents. So it has three agents that it typically would sign off on. It would be the domain and adapter agent, UI and docs, UA and security. So UI is user interface, and then UA is quality assurance.

Within these they have ownership and what they own in this instance is domain and adapter. Hold on, I have to get some water.

Dalton Anderson (50:43.928) All right. So the domain and adapter agent, it owns the database, the Docker instance, all that stuff. The models, the business logic, and then the UI and docs owns the components, pages and

the documentation and it's spun up if there's some kind of browser session needed. And then the U the UA or sorry, the QA and security, it it's used when there's multiple repos involved. And so it talks about how to structure a goal. And the main thing that you need to do with the goal is tell it the "what." And then you gotta tell it how should it it meet these conditions.

Because a goal will run continuously until it can't anymore. So you can tell it what to do. And then the last thing you need to tell is how to measure that what you did worked. Because if you just tell it what to do, it's like, "Make this thing." And then it'll make it. But it might not be good because it's just gonna do whatever it wants. So you've got to structure it and tell it, "Okay, this is how you'd validate that what you did

was what I'm thinking that you should have done." And then from there, you can iterate. And then there's some verification gates that have been put in place. So this is the hex validate. So hex validate is what I think is the right way to think about architectural pieces within your platform. It kind of depends on what you're doing, but for a lot of stuff I'm doing, I want a hexagonal approach, which is similar to modular, but modular is different. Like modular is like,

"Hey, I don't I want I want to have multiple services in s instead of a monolithic code." Hexagonal is saying, "Hey, not only do I want modular services, but I want inputs and outputs to be these separate entities to where I could switch out something that I didn't want before and I can plug and play technology." So if I wanted to change my infrastructure from this this API to

Dalton Anderson (53:00.154) whole different API, that's fine because it's not integrated anywhere else. It's just in that one spot. That's a really bad example, but it's really easy to switch technologies instead of if it wasn't like that, it's a lot more work to set it up hexagonally, but it provides a rigid structure for AI agents to follow, which is very important. Cause if if if it's not rigid, then it's hard to verify that it's working.

And then it can spire out of control.

And then there's security audit. So this is these audits. Like you don't want any of your secrets. You you don't want any of these like leaks to happen within your platform when you're posting stuff on your repos. And then there's the UI slash like lint. And so it just reviews the code. And then there's some auto verifications. So I have built-in auto verifications for Golang, for Rust, and then for TypeScript.

Dalton Anderson (54:06.67) And then the setup, this is the most important part. You would just copy and paste these things into your terminal, which is the terminal commands. And then you would just be on on the go. One thing I might have to do is you might have to put like what to what to put there. But anyways, this is my first attempt at providing this information in this way. I hope that it's super useful

and that it pr provides a better outcome for you. There's a lot of extra stuff that's built in here. Like one of the most important things that I probably didn't touch on is if the agents are ran, they're only ran at runtime. So what that means is if you start a project, you don't have all your folks around you,

you know, if you're starting a project and you need accounting and you need some technology folk, you don't have all these people gathered around the water hole, watering hole, waiting for you. They're doing their own thing. They're working. And then you call them over when you need them. And so this instance is you only wake these agents up when you need them at that moment. Sometimes if you don't have it set up that way, if

it knows it needs that agent, it will just turn it on and then you feeds it con contexts and then from there it's wasting context and then it's just inefficient. So you want to make sure that you're not wasting your tokens and that's one way to prevent

Dalton Anderson (55:51.132) There's many other instances. And then just to make sure that this is clear is this these verification gates, these are custom commands that you can run. They won't auto fill, but if you type them in, they will pop up. But yeah. The token efficiency right here. Subagents spawn on demand by default is off. And then max three concurrent directors and two concurrent specialists. Subagents auto kill on task completion. And then parallel rights use branch isol isolation.

And then tasks completable in under five hundred lines of output, always inline, never sub agent.

So those are the rules to prevent this ballooning of token usage. Cause you'll see stuff online. They're like, "I'm running 50 agents and I'm getting so much work done." But really, what they're doing is super inefficient because that agent needs to communicate with the other agent and that agent needs to communicate back. And that agent communicates to another agent. And then before you know it, it's the classic game of telephone. And information might change, but also it it's not efficient. So

you want to prevent that when you can, for the most part.

Dalton Anderson (57:12.197) That was Antigravity and Antigravity Command Line and Antigravity 2.0. I've had such a great time building with these tools. over

Dalton Anderson (57:29.915) Think Antigravity 1.0 was released early this year or late last year or something like that. I've been having such a great time with Antigravity. I love it. I love Antigravity 2.0. I love how generous their credits are if you don't have one of those crazy subscriptions. And I love how easy it is to build stuff. It's great. I'm a

huge fan of Google and their products and this is another one that I'm fond of. Hopefully it doesn't go on the Google kill list. Who knows? We'll see. But I doubt it. This is really important to their future initiatives and it's very clear that they think so as well. Of course, wherever you are in this world, good afternoon, good evening, good morning. Thanks for listening and listen in next week.