Google Antigravity Livestream: Behind the Product and Gemini 4 Argon

Author: Google Antigravity · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Google Antigravity Livestream: Behind the Product and Gemini 4 Argon · Channel: Google Antigravity · 37:08 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, Prompt and context engineering, AI developer tools, AI models

Summary

The livestream reviews recent Antigravity launches and product changes, including the Gemini desktop app, Google Books, a redesigned right-side panel, and expanded customization and plugin features. The guests discuss how increasingly capable models can use terminal commands, delegate tasks, and work with less detailed prompting, while emphasizing the need for predictable, safe, and verifiable behavior. They compare Gemini 4 Argon with Flash and share how model capability is changing coding workflows, context management, browser use, and automation. The conversation closes with underused features and ways the community can share feedback and plugins.

Key points

Recent launches discussed include the Gemini desktop app, Antigravity on Google Books, and Command-K advanced search.
The right-side panel is being rethought to accommodate conversations, artifacts, sub-agents, and other supporting tools as the product expands.
A new customizations area aims to make plugins and Google tool integrations easier to find, install, and manage.
As models improve, they can often use terminal commands directly instead of relying on thin MCPs or other shortcuts; the product still needs to make this behavior safe and understandable.
The guests describe Gemini 4 Argon as better suited to ambitious, multi-part tasks and delegation, while Flash offers fast feedback for iterative work.
They recommend adjusting the level of human involvement to the task: delegate when results are easy to verify, and stay closer to implementation when nuance or security matters.
Long-running conversations can preserve useful context across related work, but it can make sense to archive them when switching tasks; very large context windows also bring costs and quality trade-offs.
The team times product features to model capabilities, using internal and external feedback to assess when a feature is ready.
Voice input, sidecars, generative UI, terminal access, and remote control are highlighted as useful or underrated features.
Users can provide feedback through social channels, forums, in-app feedback, and by sharing plugins or customizations with the community.

Chapters

Start	Chapter	Summary
02:30	Recent launches and the evolving right-side panel	The hosts review recent launches and discuss how the right-side panel must scale as Antigravity adds more content and interactions.
05:03	Customizations, plugins, and model shortcuts	The guests introduce the customizations area and explain why stronger models may no longer need some previously useful tools or shortcuts.
07:05	Terminal use, safety, and verification	They explain why models often use terminal commands, how that can be efficient, and why agent behavior should remain predictable, safe, and verifiable.
11:45	Advanced search and the guests’ roles	After mentioning Command-K search, the guests introduce their work on Antigravity and the value of community feedback.
13:48	Gemini 4 Argon and changing expectations	The guests share their early experience with Argon and describe how improved capabilities have made agent-assisted work feel more collaborative.
15:54	Choosing between delegation and iteration	They contrast Argon’s ability to work toward broad outcomes with Flash’s usefulness for fast iteration and tasks requiring closer human judgment.
21:05	Ambitious workflows, voice, and sidecars	The discussion turns to future product ambitions and features such as voice input and sidecars that can automate recurring work.
23:39	Prompting and managing long conversations	The guests describe using fewer, longer conversations for related work, separating tasks into distinct streams, and archiving when changing focus.
27:19	Context windows and the browser	They discuss the cost and quality trade-offs of using very large context windows and how agents are evolving beyond browser automation infrastructure.
29:22	Product decisions and model choice	The guests explain how model readiness and internal testing help determine when features should launch, then discuss when to use Argon or Flash.
33:03	Underrated features and remote workflows	They highlight generative UI, terminal access, and remote control, describing how agents can work in the background across devices.
35:36	Sharing feedback and closing	The guests recommend sharing feedback through community channels and publishing useful plugins, and thank viewers for helping guide the product.

Quotes

don’t get in the way of the model — 07:05

Trust but verify. — 11:45

I think you should just be ambitious with what you want anti-gravity to do. — 21:05

while it can it doesn’t necessarily mean that it should — 27:19

it just gets work done for you. — 32:32

Concepts and entities

Google Antigravity, Gemini desktop app, customizations, Google Drive, Chrome DevTools, Gemini 4 Argon, Flash, Kevin, Rodie Devril, Google, CDP, Playwright, generative UI, Remote control, Google I/O

Full transcript (original language)

02:30 All right, we’re back and live and super excited to chat with you guys again. It’s been only a week this time since we last were on the live stream and we got a cool guest here. Hey everyone. Um, so we uh just wanted to kind of quickly go over some of the things that we launched

03:00 since the last time we did the live stream. Um, first of all, we launched the Gemini desktop app. Uh, did you get a chance to use this? >> Yeah, it’s pretty cool. >> Yeah, it’s a I I got to try it out and it’s just like a simple toggle. My favorite is the command shift A, I think it is, to switch. >> Is that the hotkey? >> Yeah. Yeah. Yeah. So, it’s um I I know it’s not maybe fully featured. You know, there may be some stuff missing, but I personally love like the Gemini desktop app. It’s got the meat of, you know, most things that you would need. >> Yeah. Yeah, absolutely. >> Also, it’s convenient. It’s just, you know, >> Yeah, for sure. Um, we also launched on

03:31 Google Books, too. Um, >> that was a fun launch. >> Yeah. Yeah. Um, I I think it’s really cool seeing anti-gravity starting to show up in a lot of different places. So, >> uh, yeah. Um, we also kind of been rethinking the ox panel. Do you have any thoughts on like kind of how that’s been transitioning and new design for that? >> Yeah. Yeah, the ox panel has been one of those things that in the very beginning we had we were in the IDE world, right? And so over time the ox panel has become more and more loaded with things that we have to visualize, things that the user has to interact with.

04:02 >> It’s becoming more of a two-way street, window management, pain management. So we have to rethink a lot of these patterns to just make them scalable for this concept of, you know, a super app where you’re going to spend all your time inside of this interface. >> The right panel just needs to grow up a bit. >> Yeah. I think one of the things that people uh had either concerns or like questions on sometimes things will get hidden like that’s where you saw sub agents running and stuff like that. So what’s the new is it is the ox panel the one that shows up when you have the sidebar collapsed. >> So the ox panel I guess from a technical perspective we defined it as anything

04:33 that sits in that >> right side if you’re facing your screen on that right side. So conversation will happen, artifacts will appear, sub agents will appear, anything that is sort of call it secondary to the conversation or supporting evidence or or terminal commands, you know, supporting supporting uh artifacts, all those things will appear on the right side. Um and more and more things are appearing. >> Yeah. Fair. Yeah. Yeah. Totally. Uh you want to um show the other thing that we did too >> um is we showed off the customizations tab. This is like a whole new big

05:03 section. >> Yeah, we can get in here. Um, so yeah, I I was pretty excited to see um this kind of show up because, you know, things were kind of getting hidden and like your settings dialogue, everything felt like it was going there. So >> yeah, we’re really making a big push on customizations, plugins generally. Uh I think if you zoom out and think about it from an engineering perspective, the agent is quite capable of doing a lot of things. I think we’ve seen this, we’ve seen this at Google internally with the version of anti-gravity that we have here. We’ve seen this externally with people configuring their product in all

05:33 these different ways and >> the skill barrier call it you got to go into you know what is it your Gemini config you got to like get in the terminal you got >> got to be a developer >> you got to be a developer right and the definition of a builder everyone wants to get in and customize their product so we really want to make this a great experience and so the first step of this is all right we you know we’re a Google product we’ve got these incredible suite of of Google tools let’s put these in a place where users can one click authenticate, use and jump in and just

06:04 now have your agent have access to things like Google Drive or your Chrome Dev Tools. And so super excited for this to take off. There’s a lot more room to mature and grow, but >> pretty excited that we have a good starting point. >> I also like that. Um, well, one, this is the build with Google plugins. So like they kind of were there but like very hidden. And then also we have like an installed tab that like >> shows across your system where everything is. And I think that that’s always nice cuz like sometimes you’re like how much context is being taken up by the different things. >> Totally. And very easily, you know, I I’m victim to this. I’ll throw something

06:34 into my into my plugins into my customizations and I’ll forget about it and I’ll I won’t know until I get it installed and now everything’s in one place. >> We’re going to have more tips later, but maybe uh just kind of off the bat since we’re on the screen. Do you have any thoughts around like when should we be evaluating when to remove MCPS or different types of skills that may have been needed for maybe small models but not as bigger ones? >> Yeah. One of the biggest learnings that I’ve had over the years is I mean we’ve talked about scale with intelligence. I’m not sure if we talked about it on this podcast specifically, but this lesson of don’t get in the way of the

07:05 model just keeps coming up again and again and again. And I’ve found myself, you know, I have things installed that maybe would guide the model towards particular >> functionality that now could just be terminal commands. Like the agent is so good at at using the terminal at running commands that >> frankly like an MCP that is fairly thin, call it uh maybe it’s just >> I don’t know a debugger or it starts some processes and it was almost a >> a shortcut for the model. the model is now just so good it doesn’t need those

07:36 shortcuts anymore. It can just sort of do them. >> Maybe one thing that like the community’s brought up multiple times is uh some people think that it’s a failure mode when it starts to not use the built-in write file tools or whatever and it’s running Python. Could you maybe like explain more about like maybe how that like saves tokens or like why that period? >> Totally. So if we think about this from a very basic like training perspective, the distribution of data that we feed to the model, think of all the places that terminal commands show up in blog posts

08:07 and in textbooks and you know the vat of information that’s available online >> and how long it’s been used >> and how long it’s been used for. That is what the model knows very very intimately. It is it is baked in the very core of the model. Now then you can add layers on top of this. You know, one example is you add the ability for it to rationalize what slash text is as a skill. Sure. >> Um you add little bits of data that tell it, hey, if you see an MCP, this is what you do, right? And you can imbue these these behaviors into the model. But what we’re realizing is over time more and

08:38 more these models are getting so smart that that foundation is actually what >> drives the progress forward and can do the most. And so >> um when we think about okay, it doesn’t use the right file. You know, part of that is called a product thing. You know, a failure of if it’s catting a file on the terminal, it’s equivalent to reading the file. And we should just wrap that up and make it a bit more predictable to the user. But secondly, it can do really sophisticated things with terminal commands where it might read bits and pieces of the text in

09:10 specific areas that just >> the range ones. Yeah, totally. >> Yeah. Or custom tool, our custom read file tool might just not do. It’s been very common on Linux especially to like you know do a tail and just like get the last couple error logs of this massive >> Yeah. Like the classic example is okay if I only gave you the ability to read one page at a time of a book you would and you had to find a particular snippet you would read probably front to back the entire book and find all the snippets. Instead >> with a terminal >> you can not only read the snippets that come back but you can search right and so it’ll do the most efficient thing to

09:41 get the job done. Now, it is our job to make sure that that model behavior is both predictable, safely done, and visible to the user. And obviously, there’s a never- ending list of things that we could do there. But that is generally the direction agents are going in. >> And I think maybe along that same lines, but um one of the things that I I think maybe uh threw me off at the first, but now is fine. But like people are like, well, it’s using a language that I wouldn’t use. So like, you know, I did a bunch of TypeScript or Dart and it was like it’s writing Python. It’s like I didn’t write Python. So it’s like, yeah. any thoughts on that of like

10:12 >> Oh yeah, many thoughts on this. I mean it’s it’s now >> I’m questioning like what it means to actually understand a particular coding language because now I’m like yeah wrote some >> random script somewhere and it’s I’m sort of trusting that it’ll execute. Um I I think more and more you obviously we’ll put up guardrails. I’m actually quite excited for some of the safety features that we’re going to have coming out around >> sandboxing and and sort of more intelligent uh ways to keep the model aligned. Um, but even without that, the models are just becoming so good. If you

10:43 look at, you know, Argon, I trust most of the scripts that come out. It used to be the case, call it a year ago, you would be proofreading those scripts to make sure it’s just not going off the rails. >> Oh, for sure. I think maybe even part of it’s due to like some of the benchmarks are showing has like a really low hallucination rate, you know, which, you know, obviously it’s not zero, but like I feel like I’ve noticed that so much more where um yeah, I like it can do a lot more that I’m trusting and I’m able to go a lot farther because of that. So, it’s just raising the ceiling. >> Totally. And like any other call it

11:13 security policy, there’s often layers to this. And one of the layers is yeah, you could go through and read every line of the code or >> call it a model that is outside of the loop that summarizes what the script does. >> Call that a 90 plus% >> non-h hallucination. I mean probably more than that. You read that. Okay. Now you start layering these like pretty reliable but not bulletproof systems and you get to a point where you can start trusting more and more. I think maybe one area I noticed it in the short term is like when the hallucination rate was a potential to be higher.

11:45 >> I felt like I was writing a lot more tests to like catch every all the times that it didn’t fail, you know, and so like maybe now I’m able to go farther with its output and then like have maybe more spot checking, you know, stuff like that. >> Yeah. Trust but verify. >> Exactly. Totally. Yeah. Um and then one maybe one final feature that we added since then is uh command K for advanced search. >> Oh, I love command K. >> Yeah. So, um, for people that don’t know, uh, Cameron, who was on the live stream last time, has amazing shortcuts for everything and, uh, that was one we learned. Yeah. Yeah. Totally. >> Um, okay. So, um, as we get this live

12:16 stream, can you tell us a little bit about yourself and who you are? >> Sure. Yeah. So, my name is Kevin. Uh, I’ve had the pleasure of of representing anti-gravity for for some time. Uh, I have been working on AI dev tools for the past oh man three, four, five years. Um, and it’s just seen products like, you know, products like anti-gravity explode over time. Um, I’ve been working on everything from, you know, the harness to the front end to whatever sort of makes the product successful. Uh, I now lead part of the product team

12:47 to just >> get anti-gravity out to people, get anti-gravity out to Googlers internally and just really maximize >> the model can do one thing. How do we deliver that to people? >> Totally cool. Yeah, that’s awesome. And uh I’m Rodie Devril on antgravity. And so oftentimes I’m able to work with you guys and bring what you guys are making and uh bring it to developers and all types of users. So >> it’s been it’s been great to have Rodie here because he can tell us exactly what you all are thinking and that has been critical. I mean a lot of features have come out of just >> totally >> during organic feedback and I can’t read

13:18 everything. Engineers, you [clears throat] know, we all can’t read everything and somehow you you were able to boil the ocean and figure this out for us. >> Yeah, it’s it’s pretty wild. Um, we we’re really uh receptive and uh thankful for our community. There’s a lot of really cool stuff. Um, often, [clears throat] you know, we we’re’ll talk about it more later, but you know, Googlers uh use our product and so it’s been interesting when we have a feature that goes out and Googlers have a very specific way of using it and then users first day are like, well, that’s not how we expected to work. And so, um, it’s been interesting. Yeah. >> Playing two sides of the coin. Yeah.

13:48 >> Yeah. Yeah. Totally. Yeah. So, uh, one of the big announcements that we just made is that, uh, Gemini 4 Argon is available in Ante for Ultra users. Uh, have you got a chance to use it? And what are your thoughts so far? >> Uh, I think we can publicly say, you know, we’ve been using it for some time. >> Um, >> absolutely. >> It’s been a joy to use. I mean, I I my expectations have drastically changed. Obviously, we’ve seen different variations of of Gemini and and one of the key parts about, you know, making sure our team is building the right things is understanding where models are going and seeing those early checkpoints

14:20 is really a good way of saying, you know, oh, this capability or we can expect this amount of token usage or the efficiency and tendency of the model goes in this direction. Um, I think you see a lot of this with I don’t call it the multimodal nature of the model. you’re starting to expand the breadth of what the agent is capable of doing. Uh a year ago, I probably wouldn’t have felt comfortable putting out a Google Docs plugin because just >> the general modality of Google Docs wasn’t something that the models intuitively understood. Yeah. >> With Gemini 4, these things have

14:51 completely changed and now I’m I’m writing a lot of docs with >> Oh, yeah. >> agents and whatnot. So, it’s been cool to see and and use and feel and develop an intuition over Gemini 4 and I I hope that people are enjoying it >> publicly >> and uh would love to hear more feedback. Please let us know about like you know what what you’re seeing work great, what things it could be improved on. For sure. >> Um I think one of the exciting things about you know Argon is like I I love using Flash because I felt like I could put all the rails on and do just so much work with it. But with Argon, it’s like

15:23 I I’ve been like rebuilding apps that I worked on for 2 years in a day. Like it is just >> it’s humbling. >> Yeah, it is. [laughter] And I just like I didn’t expect that. Well, one there’s like [clears throat] it’s technically capable of doing something, right? But it’s like >> it’s doing it at a level where it’s like, you know, I’m able to like collaborate with it a lot more. It’s like my intent like, you know, creating docs and stuff. Whereas before it felt like a model was just kind of like outputting an artifact, now it feels like I’m working and it’s getting much more of my intent across.

15:54 >> Totally. One of the big, if I had to characterize >> a broad generalization about maybe the jump between the flash and now Argon um flash I think we were just getting to outcomes >> faster. It’s a faster model and you were very much in a companion loop where it would produce an implementation plan. and you would read that and you’d spend a lot of time crafting that plan >> and the models were already good enough where you could say all right if my plan is good the the outcome should be pretty good. >> We’re now almost getting to the point where we talk about inner loop outer loop like >> you’re just sort of specifying what you want your end goal to be and the model

16:27 will just sort of figure out a lot of the details and you can sort of trust that hey if I want to rebuild this game and it’s capable of playing said game and and looking at it screenshotting it >> you’re pretty confident that it’ll get to the right outcome. >> Yeah. Yeah. Yeah, for sure. Like maybe uh speaking of that, like um how do you know when to use certain kind of slash commands and when we should like you know let the model take the take the reign so speak and also just knowing that like um obviously if you’re doing a hybrid like you know argon versus flash

16:58 thing it might be even a different thing there too. >> Totally. There are certain this is hard to vocalize the intuition that I’ve built. Yeah. I think a lot of it just comes from using these things day in and day out and feeling like certain types of tasks call it uh I don’t know you’re making some style changes to a front end. If I feel like the outcome is pretty verifiable by the agent >> I feel much more comfortable >> not being in a pair programming state and being in more of a delegation state um using sub agents or whatever to just

17:29 say all right like go figure it out and I’m just going to verify >> that your verification was indeed correct. Yeah. >> Um >> there are other classes of problems where I’m like, oh, actually the nitty-gritty implementation detail. There’s probably five different avenues to get to this end outcome. And it takes a little bit more human nuance to understand, hey, you know, this stakeholder wants this. >> This customer would want this. There’s a security implication here. It’s my job to to >> to >> hold that context and and direct accordingly. And that’s when I think I’ll reach for more of a flash model

18:00 because I’m iterating very quickly with said agent. >> Yeah. I feel like um with Argon especially, I feel like I just give it like five different things in a prompt and it like it’ll go off and do them all. But also, it’s probably one of the first models that I have felt comfortable running overnight, you know, not with um you know, teamwork or anything else. Like it just like >> I know that when I check on it in 3 hours, it’s going to be done like and it’s going to have all the stuff and >> it might need some input, but it’s like >> as far as the thing that’s being done, it’ll be

18:30 >> Yeah. I mean we can go down a deep dive on the actual implications of what this this means but one of my fun scale with intelligence stories. So I I helped work on some of the multi- aent orchestration work that we showed at Google IO uh not too many months ago and we we rebuilt Doom that was >> yeah right >> that was remarkable to me. I mean, we were using that was with teamwork and the idea was, you know, we’re putting up >> random injection points, random guardrails, uh, saying, “Keep going, keep going. Here’s how you run the thing. Screenshot yourself.” Like, “Please, please screenshot, verify,

19:02 verify.” And >> use 10 agents and like you’re going to have five different front-end agents and five different and now all of a sudden these these guardrails are just slipping away because it’s it’s becoming >> uh, >> I don’t want I almost use the word human nature. That’s that’s the wrong word. It’s >> becoming part of the model nature to >> to set up systems like this and that’s where you’re seeing some of these [clears throat] gains where you’re comfortable >> with it overnight. That that is just >> it’s mindblowing. >> Yeah. Cuz sometimes you need that let’s say bureaucracy, you know, like you’re adding that to catch things along the

19:33 way cuz during that handoff usually it’s like oh well I did three out of the five or you had another person. >> But I’m noticing that like you know it it’s just more capable of outputting so much more. I mean I know the model has a much longer output context and like he can just it’s more concise I think that too whereas like flash would you know generate something generate something generate something but like argon’s able to like oh I can see how this works across this like let me come up with a unified abstraction like and so I’ve been really enjoying that >> yeah this is where you know benchmarks and actual usage don’t necessarily align

20:04 and I think we’ve cracked something here in the last 3 four months at GDM where yeah you know we want to show positive benchmarks scores. Uh, but also the the presence of anti-gravity inside of Google has really made it possible for us to to AB test very effectively, very efficiently. You got 200,000 I don’t know how many people work at Google. Someone fact check you got a lot of people using this product every day relying on it 8 10 12 hours a day and you can get a lot of uh feedback on just totally this model has this tendency.

20:35 You can look at the data, you can look at qualitative feedback, and so yes, we’re looking externally, we’re looking internally, and both of those things let us get a a good iteration loop with our customers. >> Yeah, that’s that’s awesome. Um, so maybe any final thoughts on that like of what does this mean for anti-gravity’s future like Argon? >> I think you should just be ambitious with what you want anti-gravity to do. Um, I think we will roll out features in the coming weeks, months that will hopefully encourage that level of ambition and and make it so that

21:05 building these large scale projects or doing large scale refactors, working on complex code bases, uh, doing complex tasks will become easier and easier. Um, it’s remarkable when you think about what people are doing internally, just sort of hacking together. We want to bring that sort of power user experience to everyone. But what you can do for yourself now is just sort of question like what am I doing that maybe is holding the agent back or maybe unnecessary prompting or >> an an automation that maybe I’m >> overengineering and and I can just use

21:37 English language and maybe use my voice now that we have a great voice model like there are just you can expect more out of the product. >> Mhm. Yeah, absolutely. So maybe since you work a lot on thinking about the product, what is your favorite feature uh that we have? >> Favorite feature? Shoot. just I just mentioned voice and it would be hard to hard to pull that one back. I I love >> the voice modality is interesting because while I’m at the office I’m not necessarily speaking to my computer because that would probably bug the people like really quiet [laughter] the AirPods in and you know but but the

22:07 ability for >> how how do I put this concisely? It used to be the case that you had to be very very precise with your language in order for the model to do >> only the thing that you wanted it to do. Now, with the introduction of both these voice models and the call it brain of anti-gravity, um you can sort of word dump. You can just word vomit and and as long as there’s nuggets of pieces of information, the general gist of the thing will get done. And I think that’s a really powerful way of like

22:38 >> just reducing input friction. So, I’ve been really loving voice >> plugins is also something that I’ve been really enjoying. Actually, sidecars I’ve started getting into. I recognize it’s a bit more of a power user feature. So for the for the power user devs out there, like continue to lean in. We’re gonna work on this part. But I now have these random automations going off in my life that >> everything from consumerry, you know, check the price of X every whatever to every time someone tags me to PR like go off and do this analysis so that I can come in and

23:08 >> just be ready, be prepared for that event. Um there’s a lot of cool stuff there. So I I use um Sidecars for I have a server and I have it run a health check every day and if there’s an issue it can automatically bring something back online and it’s like that’s where I want an agent so I like can’t do it. Yeah. >> It’s insane. >> So great. Yeah. Yeah. So um is there maybe ways that you’ve changed your prompting style from like earlier versions of anti-gravity to now? Cuz I feel like you know the models are always changing. I felt like before I used to be very explicit like you said, but but

23:39 I think also maybe um even down to like how I’m using skills. I I feel like I can just like I’m getting into this mode where I’m like I want to just get work done >> and I’m feel like it’s getting a lot easier to do a lot of things. >> Yeah. Maybe the biggest change in my behavior and shout out some of the folks on on our team who have pioneered this and made it a bit easier to do. So I I’ve started using I don’t want to say a few fewer longer conversations. >> Oh true. >> And that behavior

24:10 >> has really unlocked I mean there’s a bunch of nice properties memory. Um the context length has gotten the context handling has gotten much better to the point where you can accumulate these quite long conversations and it’s mentally helped me stay organized. Uh, so I’ll have different swim lanes in my life, whether that be, you know, I’ve got an agent that helps me with management, I’ve got an agent that helps me with, um, you know, uh, backend changes, another for front-end changes, another for harness changes, and being able to keep these all streamlined into

24:40 their own separate buckets. I’ve, this is a new behavior that I’ve found and mostly powered by the fact that >> the newer models are so much better at doing self-chestration. Yes, >> it’s no longer the case where every single step prior was just going to pollute your context so that your next task is completely different. Now it’s a bit more like oh if it’s going to be a big enough task it’ll send it to a sub agent. It’ll send it to a different conversation and [snorts] >> actually this >> task was you know not that many tokens in summary in that main stream. So that

25:12 has been a way that I’ve changed my work behavior and prompting. >> Yeah. I mean, I feel like um I always had to be very explicit to be like use sub agents for this or like delegate with this command. Like um I I definitely noticed on the newer stuff it’s more natural. Yeah. >> Yeah, for sure. Um is there any maybe along that line of like any context tips on like how you’re managing these long conversations and what to do for I mean I know people will it’ll either get slower if it’s really long or you have like a lot of artifacts like when when

25:43 do I need to like create a new conversation because I feel like that’s a [laughter] question. >> I don’t know if this is a necessarily a solved a solved problem. I think for me >> I have in my head certain >> feature milestones that I’ll there’s a bunch of features that I’ll group together uh you know we might think something I was doing internally. It’s like I was trying to bring Google Docs to uh anti-gravity and it was like okay there’s >> probably 10 different PRs that go into this. I’ll just keep the same conversation for the duration of those 10. It’ll accumulate that context. any docs that I was linking in or markdown

26:15 implementation plans that we just have in memory. But then it was like, okay, I’m moving on to a completely new thing now. I’ll just go ahead and archive that and and switch over. Uh I don’t have an amazing answer for this because I have noticed now, you know, my artifact list is getting quite long. There’s different things that we are looking into right now to make this a more pleasant experience. But um >> not a science, but I when I completely switch tasks, I think I would generally >> Yeah. Yeah. I feel like for me it’s like um when I finished doing the thing I’m like sometimes I’ll do slash learn but I

26:47 just like I like thinking of I start it I do the thing I close it I archive and I just kind of go so I’m basically I have open issues and I’ve been trying to tie a lot of my work to like >> issues on GitHub or something like that >> mostly so I can have an artifact of like you know a markdown file with a list of the code that was actually changed and so it can all kind of be >> totally >> colllocated. Totally. Um maybe something that like you know our community is like always asked about is like well if Gemini has a 1 million input and now output context window like why do harnesses not utilize the full context?

27:19 >> I think you know while it can it doesn’t necessarily mean that it should and the reason why is because of things like latency or just uh the the quality of dealing with that many tokens. It’s both it’s very expensive. >> True. >> And so you know we try to find that balance. So, it’s slower, it costs more, and it isn’t necessarily, it’s not perfect memory, right? Just because I have a higher capacity to remember, >> you know, >> if I watch seven movies and I can remember all seven versus just one

27:49 movie, I’ll probably remember details about the one much better than I will the seven. That’s true. >> And in the same way, like I think models, it’s just not perfect context and there is some degradation over over time. And then the other thing that maybe um as it’s changed too, a lot of people um you know especially early anti-gravity users use the browser like is there how are we thinking about you know just the browser evolving in general because I feel like at first it was like wow I can control the browser and that was such a big deal but now it’s uh it’s been interesting to see that evolve over time. >> Yeah, we’ll have a lot more exciting

28:19 news coming for the browser soon. I think the way I’ve thought about it is, you know, previously the only way for the agent, again, those guardrails, the only way for the agent to verify that, you know, your web app or your game had worked was by spinning up a full-on Chrome system and >> running an extension. And it it was very cool. It was mind-blowing for the time. Now you’re starting to see that it’s pretty good at using open source libraries and systems like CDP or uh playright you know and it doesn’t need

28:50 to reach for the heavy infrastructure that we put in front of it. >> So we’re doing some work here because I still think the core premise remains the the core idea is you have context and information in your browser and you have actions effectively just buttons you need to click >> inside of a browser. >> The agent is able to now >> do this reasonably well. So, we’re trying to bridge that gap and I think we’ll have something cool to show and and >> awesome. Yeah. And then um maybe also working on like a product like anti-gravity. I feel like there it comes

29:22 with tons of opinions like we have everyone across the board including community. It’s like you know like I wish I did this or like this actually doesn’t work and then you have like conflicting stuff. How do you go about like you know thinking about what to prioritize and like maybe the answer is like when you know as opposed to if it should or not. Totally. The [snorts] this industry is now quite funny because ideas are becoming cheaper and cheaper because the experimentation is so easy. And so, you know, we might have had ideas >> two years ago that are valid, right? This idea of uh what’s a new one like I

29:54 don’t know, plugins has been around for a long time. Maybe that’s not a good example, but like orchestration >> maybe that’s a good >> and just over time the it just wasn’t ready at the time. The models weren’t good enough. And there is a heavy amount of living close to the model and understanding where the model is going that will then dictate all right in the in the road map this is probably something that we should do that will land around this time where the model will also be able to now saturate that part of the product because if you release something too early

30:26 >> you won’t actually reap the benefits because you built all this heavy infrastructure and yeah you’re compensating for model weaknesses. you release too late, then you’re just late and there’s a model overhang [clears throat] and we don’t want that. Uh so you sort of have to time it right where the the intelligence for that feature is just right at the the the right moment and that is a challenging thing and this is why having internal feedback and iterating very quickly on the internal version of anti-gravity helps to then understand okay for external anti-gravity this the time is now the time is >> yeah because I feel like sometimes we maybe I there’s features that I feel

30:56 like we had internally for like six months you know before we released externally and it’s >> you know I feel like we have sometimes different use cases where maybe people weren’t running into the same stuff and but then there was a moment that I felt like it becomes useful. >> Totally. Yeah. I mean we’re in the business of giving the best most capable frontier intelligence to our users and sometimes models just need to catch up a little bit. >> For sure. Yeah. Speaking of like um so obviously with people um having certain amount of quota and capacity for for their plans, how would you go about like

31:29 choosing when to go argon or flash or maybe a combination? Yeah, I think it come it sort of comes down to that way of thinking that we were talking about before. There are certain tasks that are just set up for success in the argon world. Uh and then flash. I use flash on I don’t want to say most. I actually don’t know what percentage I’m I’m reaching for flash, but I’m using flash quite a bit. That >> immediate feedback is unparalleled. You can’t get that with a different model. And so >> there is a bit, you know, I’ve heard the schools of thought around, okay, you come up with a plan and that plan takes

32:00 a lot of research. M >> okay now go off and execute using a smaller model like flash you can do those things um I’ve also seen uh maybe the overnight stuff being something where speed and you can afford slower speeds >> uh so yeah different different intuitions >> totally okay that’s fair yeah I feel like um >> yeah it’s been interesting to be like okay even in my like prompting I’m like I’m not going to waste an argon one on this [laughter] so it’s like flash is

32:32 In the fullness of time, I’m I’m really hoping that, you know, models will continue to to get faster, better, cheaper, and users will have to think about this less and less. And the direction that I would like the product to go is >> it just gets work done for you. You don’t really have to worry about the minute details of what project, what permissions setup you have, what model you’re selecting. There are a lot of things that I think will vanish into the background >> soon. We’ve already seen this, right? Suddenly, we’re running multiple conversations at once. Never would have thought that was the case like a year and a half ago. Totally.

33:03 >> So, what does what does that look like a year from now? And I think >> model selection hopefully >> for sure >> maybe is one of those things. >> Yeah. Um yeah, that’s really great. Um maybe any um last minute underrated features that you feel like uh could get more attention. >> Last minute underrated features. Um generative UI is very cool. >> It is really cool. >> I love generative UI. I have a few automations set up so that I can crawl Twitter slash online forums like product metrics and

33:33 and to see that in a rich way rather than just a block of text >> is such a cool Yeah. >> such a cool feature. >> I know it’s such a small one but I still love using the terminal. It’s like there’s so many times where it’s just like I’m like remoteed in with remote control or something. It’s just >> really nice. Remote control has been has been a nice thing as well. So great. >> Just to feel like things are always happening for you is nice. I feel like I’m using it like 80 90% of the time now like in in my usage of it. So and and of course the internal version is mostly remote control. >> Mostly remote control. Yeah. This is one of the reasons why we we had some conviction here.

34:04 >> Yeah. Totally. Um I feel like uh I’m starting to use it on my fold a lot more. There you go. Yeah. >> Great. Yeah. Yeah. I feel like uh the things that I and it came down to even like a choice of computer. Like it used to be I would get this like super beefy laptop with all the RAM. I’d run everything locally. But now I just have like a small MacBook Air and it like I’m just remoting control and I never thought as a developer I wouldn’t need this beefy thing and so it’s been so nice to just like kind of take it on a plane and just kind of do whatever. So >> you really feel I used to walk you know

34:35 you you’d want three screens now all of a sudden you’re just in one app and everything can happen. And I I mean I don’t know about you, but it’s like I feel like even in my personal life, it’s like I can kind of be in the moment more too because I can just like start an agent, go do something, and just like know it’s like working in the background. >> I guess there’s two schools of thought here. I was >> now that I’ve got this fold on, you know, I’m at a wedding. I’m like, >> yeah, we got to get it done. I have no excuse now. [laughter] Everything is available. So, you know, but that’s a personal problem. >> And for Yeah, exactly. for um for users

35:05 also that if they’ve installed a PWA, what’s really nice is you can get native notifications. So, I’m even getting like pinged on my watch. Oh, wow. That’s so Yeah. And and especially I feel like I did with Flash, I felt like I do it and just like immediately notified, but now with Argon, it’s so nice cuz I could be like going and washing the car or something like that and then like in 2 hours it’s like, “Oh, okay, cool. It’s done.” And that it’s like kind of feels uh the right type of feature for that. >> Superpowers. >> Yeah, exactly. Well, uh, um, maybe as we close up, what is kind of the best way to get product feedback? I feel like, you know, we kind of talked about like

35:36 there’s a lot of con conflicting incentives and things like that if people like want to have features, but, you know, oftent times people may have like legit use cases where they like want us to do it, but like how can they best like give us that feedback? Also knowing that like sometimes people are on different parts of the AI journey too, you know, like some people may be still using the ID as their daily. Yeah. >> But then maybe six months months from now they might be using a different product. So >> I I think the traditional angles of

36:07 >> I don’t know call it socials and forums and whatnot. We are there. We’re doing our best to to meet you all where you are. I think that’s a that is still a valid and helpful way to communicate. Okay, >> I think another area is um uh with this whole plug-in customization ecosystem, if you build a plugin that you feel like injects, you know, maybe your own person like your own version of doing things and you can sort of codify this into a plug-in and wrap it up nicely and put it on GitHub or some other um Oh yeah, sharing platform and that is a really

36:37 nice way to socialize and sort of validate your idea with other folks who are using the product. Um, I found that to be a pretty effective way to see, oh, you know, >> then someone else can just instantly clone it, try it, and you like, oh, yeah, I I see what you mean. This is pretty powerful. >> I feel like because everything’s so cheap, you know, it ends up being something similar to like natural selection where eventually enough people are using something where it kind of the good stuff rises up and so then we’ll we can get eyes on it, too. >> And I know some people ask like, do we see like the submit feedback button in the app? And [laughter] we we definitely

37:08 are looking at all that. So, thank you so much. Please continue to do that. It’s like the only way is the team can even like be able to deep dive. >> Yeah. Yeah. It’s super helpful. I mean all the feedback again like we rely on this volume in order to understand what we should do with the product next, where we should take the models next, where are things failing, etc. So please keep that up and thank you if you already do this. >> Awesome. Well, uh thanks so much for joining us today in the live stream. Thanks for having me. >> Yeah, we’ll uh see you guys on socials and in the next one. Thanks.