Name: What We Shipped: Feature Updates, Tips, and Live Q&A with the Claude Code Team
Uploaded: 2026-04-07T20:31:43.847Z
Duration: 1 h 20 s
Description: What We Shipped: Feature Updates, Tips, and Live Q&A with the Claude Code Team

Transcript for "What We Shipped: Feature Updates, Tips, and Live Q&A with the Claude Code Team": Hello. Hello. Hey. I've gone. Hi, everyone. Welcome, everyone. We're so excited to have you with us. And I am very, very excited to be here with Adam and Thariq from the Claude Code Code team. I am Casey. I lead marketing for Claude Code Code, and I wanted to just do a quick, run through of housekeeping. This will be recorded. It will be sent to you along with, we will collect as many unanswered questions as possible and try and get answers and resources in your hand. There are some docs in the docs section. You can click through our weekly digest, of the change log, and you can, also, if we will try our best to get to as many questions as possible, but there's a lot of good stuff that we wanna get through, and so this is gonna be very hands on. And we're gonna preview some new features, and so we're not gonna get to every question. Suggest, joining the Claude Code Code community. There are places where you can ask questions. You can also ask Claude Code your questions, but the community is a great place to get your questions answered as well. I, I will be, stepping aside and letting Tharuk and Adam take over here, but, we will we will be, doing our best to answer questions at the end. So we're not gonna do it along the way just because we got a lot of of stuff we wanna dig into, and I think everyone will be very excited about that. So with that, Thariq, I will hand off to you. Yeah. Hey, guys. I'm Tharuk. Yeah. You might recognize me from X, but, yeah, excited to chat about everything we've launched. I I know we've launched a lot this month. I'm joined by Adam who, you know, is one of the, like, original members on Claude Code. He previously was head head of engineering at Robinhood and the senior director at Meta. So he's seen some of the fastest growing companies ever, and he's, you know, now on the most the fastest growing one. So, Adam, welcome. Do you have, anything to. It's great to be here. Yeah. yeah. Super psyched to to talk about some of the stuff I've been working on. I'm gonna do a little demo. But, yeah, let's let's go over what you got, Thariq. Yeah. So really quickly, we've got, three things. We're gonna talk do a demo on agent teams revisited. You know, I I think agent teams is a feature that a lot of people were really excited about, and we released the research preview. So I think we're doing, like, a revision there now. What we shipped, we've shipped a bunch of different things. So you can see, like, we're gonna talk about things like million context windows, memory, computer views, and Claude, Claude and Chrome, Claude Code remote, asking Claude Claude for HTML. This is, like, a new thing that we've been doing a bunch of. And we'll, we'll talk about Claude Code no Flickr, which a lot of people thought was an April fool's joke. So lots to talk about, and then then we'll get to q and a. So, yeah, do we want to do you wanna start screen sharing and and tell us a little bit about maybe maybe before that, like, tell him what's motivating this. You know, like, how did agent teams start out, and and how do we get to, what problems are you trying to solve now? Yeah. That sounds awesome. So let's see. I'll I'll go I'll rewind to, like, Christmas vacation, you know, the holiday vacation we had in December. And just like everyone else, I was at home using Claude Code a ton to to try absolutely everything. I'll make a note. Somehow, we're gonna have to unshare these slides so that I can share my screen when when. we're ready. Yes. Okay. Yes. Great. But, but, yeah, I I got obsessed with the idea of putting Claude in a mode where all I could do is delegate to sub agents. Yep. So, you know, the main Claude that I was talking to couldn't do any work, Yes. and it was pretty cool. Like, I I I called it coordinator mode. There are a lot of problems with it. You know? Like, Claude has OCD. So if you give it any way to check on the sub agents or even the ability to do the work itself, it tends to, like, race the the sub agents that it delegates to. So it took a lot of tricky prompting to get it to work, but it kind of worked. And I think, you know, it was better in some cases and worse than others, but very interesting. And it got me, like, really, well, as we say, everything is pilled at Anthropic, so it got me right. pilled. That's right. I came back, like, obsessed with this idea, and I'm really wanting to bring it back to Claude Code. And, you know, you saw at the beginning of the year, we released agent teams. And agent teams, you know, is a was is still a research preview. I know a lot of people use it and love it, and there are just so many good ideas in there. The ability to start a Claude Code out of process and have it talk to another Claude Code, the ability to address sub agents by name, you know, detailed observability. These things are great. And my what I wanted was to bring this into the core of the Claude Code experience without, compromising anything in terms of how Claude Code works for you now. And given the scale we're operating at, all the different ways people use Claude Code, that's that's no easy feat. But I'm hoping today I can demo a few things that we're working on. So maybe this would be a good time to start the screen share. Yeah. Now, Sounds great. great. I'm gonna show just talk a little bit before we we start, I I think it's, worth. calling out that, like, agent teams as a research preview is a good mental model for how we ship things sometimes where, like, you know, we're this whole thing is about what we ship. Right? And sometimes when you ship things in research preview, it's because, like, we want to learn from you guys as well. Right? Like, I think all of these models are so new, how they interact with these, like, very complex systems, like, message passing between each other. Like, we can test it and drop it, but as soon as you put it out in the wild, you get, you know, a 10,000, a 100,000 times or more, usage. And so I think when you see a research preview feature, like, expect some iteration. You know, expect that, like, you know, we're going to try and land, like like, we haven't figured out the final solution. And if you're not, like, a user where you're like, yeah, I need to, like, be experimenting, tinkering with everything, it's fine to just wait for, like, you know, the the final iteration of it. Right? It's not something where you you like it. You know, you have to use agent teams to be the most productive or something. Yeah. Totally. In fact, I hope we can talk about a little bit more because I think a lot of people feel this kind of Claude anxiety. You know, Yeah. the team shifts so much, and there's so many features. You don't have to use all of them to use Claude well. And oftentimes, you you know, for the most part, most features are a replacement for copy paste of one form or another. But, you know, like, it's amazing what you can do if you can cut yourself out of that copy paste loop. So. they're good to have, but agreed that, like, you know, it's it's impossible to know how these features should work until after you ship them. And that's part of the reason why we're just committed to working so fast is is because, we need that feedback from everybody to know. Yeah. I think there are some other things we'll talk about that are in the same bid. But, yeah, let let's see the demo. Okay. Great. Alright. So, I'll emphasize that this is a, this is just a a a development branch on my on my computer. So these features are not available. We're not committing to a time frame. If you're if you're lucky or unlucky, you'll see some bugs here. So that'll convince you that it's not ready to ship. But I'm gonna be working with this project I call I call cmux. And, you know, there are a lot of, things called cmux. I I am like, oh, yeah. Have cmux. People are like, I use that too. I'm like, oh, this one. I don't think so. I'll I'll provide a link to my, repository in a second. But, Nice. what it is is like a it's like a little extension for tmux. So tmux is awesome. It's this window manager for the terminal. I use it all the time, but it's got this terrible UI, and the key bindings are also really bad. So cmux is kind of just a a way to take over tmux and make it more useful with both an interface and also better key bindings. And today, I'm gonna be working with this layout chooser. So you can see here right now, I have one pane in this layout. And one thing that's really hard in Team Box is getting getting the panes of a window to be how you want. So I have this little layout chooser, and I can pick what layout I want. In this case, you know, when I go from one to two, I just wanna split split the pane, and I can have two Claudes. And I'm gonna get to work on on Claude, using Claude now, of course, so I'm gonna start Claude. I'm using no Flickr mode here. You don't have to, but, I've come to really like it, so that's how. I'm gonna demo this. And I have a prompt ready to go here. Oh, did you usually just say? No. Same. Same. Yeah. I'm on I'm. on NoFlick as well. Yeah. Yeah. Totally. I especially if you use tmux, I think it it kinda helps a little bit. Okay. So I've got this long prompt here. I'm not gonna bore you with the details, but long story short, what I really want is when I'm when I tell Claude like, hey, Claude. I wanna see that diff, or can you show me that file? I wanted to use cmux to open the pain for me so I can I can look at it? I don't have to think about that. Another thing that's kind of a pain to do with regular tmux. Yep. Now what you'll see here is that Claude has launched three agents to do this. So this is more parallelism than you usually see from Claude. And more importantly, down at the bottom here, you see this, sub agent menu, and I can actually jump in and look at any one of these sub agents and see what it's doing. I mean, I can actually even type a steering message in the prompt if I want to. One sec. I I don't usually find that's necessary. I usually talk to the main agent, but it's kinda nice to know sometimes what what everybody's doing. Yep. And, you know, people use plan mode. Like, I think one thing that we all now know about Claude Code is that if you sit down and you're like, hey, Claude, Can you change it so x? It'll do a it'll do it, but it'll do a terrible job. You know? It. won't read enough files. It it won't pull in enough context. So this is a way to get Claude Code kind of working on, you know, building up the context. But because it uses sub agents, it can read things that are not relevant, and those things don't make it back into the main transcript. And I think another thing we've all seen is, like, managing the context is really important. Even with a million context, keeping that context short and relevant is really valuable. Now, there's a I'm gonna start another Claude here, and this other Claude Code is going to, help me work on an associated skill. So to get this to work, I need both cmux to have this functionality where it can open something in the right pane. But I need Claude Code to understand that, like, when I say open or show me, that's what I need. Now to start, I'm just gonna have Claude Code read the docs for skills. And I always do this because I use skills all the time. I make use skills a lot, but I can never remember, like, the front matter format or even where they. go. I'm a little scared of that. So, Claude Code gonna go do that this research. Now back to my first pane here, it looks like it did a bunch of research. You know, I'm not gonna read this too carefully because we're we're on a big webinar. I'm just gonna show you another feature that we have. And, you know, sub agents can work two ways. One way is, like, Claude Code can start a brand new sub agent with its own prompt and its own set of tools, and it can it can be, like, give it a whole prompt and tell you this is what I want you to do. And. this can be really useful when, like especially for, like, adversarial review, something like verification. You actually really want that clean context. However, there's another form of subagent where it's really helpful to share that context because it means that you have no information transfer loss, and you can share the cash, for for the conversation, so it's relatively cheap. So there's a new, command coming called slash fork. And the interesting thing about this is I can give this sub agent a very short directive here. I just gave fork the implementation, and it knows exactly what I'm talking about because everything that came before is included in that sub agent fork. Now this is a regular old sub agent just like, the ones we looked at before. In fact, the other one the other sub agents we looked at, I didn't tell you, but they were fork sub agents too. So I can create these. I can create these. And and they're both really nice. So, we'll let that cook a little bit back to the skill. Now, here, I'm gonna use a little dictation, and I'm gonna tell the skill, Claude Code, what I wanted to do. Hey, Claude Code. There's another Claude Code working on the same directory. Can you get in touch with it and ask it about the feature we're building? That's what the skill is gonna be for. Okay. I can send it. And now you're gonna see another new feature that we're working on, which is the ability for any Claude Code on on one computer to talk to a Claude Code on the other computer. And you can see here that Claude's, sort of message appears in the window. The rendering is a little rough. Like, see the whole message obliterated the transcript. Again, some of the things that that still need refinement. But, you can see that, you know, that other Claude was able to respond. And one thing that's nice, like, because the implementation was forked and running in the background, the main Claude was free to answer. that quest So there are a lot of reasons why maximizing your sub agent use can be really helpful, I think, controlling how context grows, keeping the main agent free. I like to chat with Claude while it's doing stuff, and and the main agent, pattern, does it. So, you know, now I can say, yeah. That's the idea. When I say, open the diff, use cmux open. Okay. Great. This this, finished with the. implementation. I'll say, don't forget to build because, Claude always forgets to build. Yeah. Even though I have that in my CLAUDE.md, I feel you people out there. Instruction is always hard for all of us. I just wanna say, not just Claude. Well, I think it said built it succeeded, it did at this time. yeah. Yeah. You can. Yeah. It. it it's you know? all caps shouting worked. Claude. Okay. Good, Great. Claude. Yeah. Let's see. So okay. It made the skill. Now, let's see if we can get the, get the the big finish for the demo. I'm gonna start. a new Claude so it loads the skin. I'll say, hey, Claude. Can you show me the diff for this branch? Hardest part is waiting for the, translation. Does. it work? I guess, I think it competes with the microphone for the webinar. Oh, true. Yeah. Yeah. Okay. Let's see what happens here. Let's see what happens. If we get lucky, it'll, like oh, Teamwork's. There. we go. Okay. Nice. That's awesome. It, that's sick. it up in. Zemux, and now I've got my GitDip. Yeah. Oh, Alright. damn. Great. So those are some of the features that we're working on for agents and sub agents. Hopefully, it inspires you. Most of this stuff actually can already be done with some of these exceptions. Some of these built in things like fork, but working hard to get these, out in the public release. Yeah. So so to sorry. If you do you wanna screen share again, I just want to, like, kind, of sure. Look over it. to. Yeah. Yeah. Let. me go. back to that. I I think, like, some things there to recap, I I think, like, one of the things we wanted to solve was the UX of seeing what sub agents were doing. Right? Like, with agent teams, I think it was, like, a little bit hard to see exactly, and you can steer them as well. Or, like, the agent teams and the sub agent implementation were different in that way. Right? And so now we've, like, unified them. Any sub agent, you can click in, view, you can steer if you want, you get this, like, nice UI. And so that's one one learning. Right? Then I think the other thing is, like, communication. So do you wanna talk a little bit about the, like, way that, like, sub agents can communicate and, like, it seems like they can communicate with each other and also with another Claude. Right? That's right. Oh, yeah. A a bunch of interesting things about this. So rewinding. a second to, like, you know, Yeah. agent teams as a research preview was. this capability that was kind of bolted on. It it. reused some of the same tool names, but a a lot of the code best were, like, unique. So. some of the work I've been doing has just been trying to go back and, like, clean things up and merge it in. I'll just give an example. At the beginning of the year, the tool for Claude to create a sub agent was called the task tool. Yep. So you'd be working with Claude, and you'd be like, hey, Claude. We need to, like, you know, make this improvement to your agent tool. And it would be like, I don't know what you're talking about. You know? Yeah. Or you post on Slack to one of your teammates and be like, I'm working on the agent tool. And they'd like, what do you need? And, you know, because it was named the task tool, and we also have this task list inside Claude code, Claude Code. would get super confused about like, it's weird because the task tool has nothing to do with task create, which is, like, for the to do tool. So, Yep. part of what need to happen here was actually rename that tool. Now I I did this, and in fact, I actually broke SDK users for a version, unfortunately. I didn't think through, like, exactly how this would show up in the SDK. So just some of the some of the stuff that goes into, like, what we're talking about when we talk about, like, kinda clean this up and really get it ready. Yep. Just, like, even building migrations for some of these ways that we need to be able to change the SDK from version to version. Yep. So thanks for that prompt. That's a great example. Now, other things that that we've seen here, there are a lot of different concepts, and I think the thing that we try to avoid is, like, giving Claude 50 jillion tools that. are all very purpose built. Like, that's a real anti pattern, you know, and it's it's pretty easy to add another tool. It's very hard to remove it. It becomes this API shape that people depend on for the SDK. And it also, you know, the other thing about lots of tools is, like, it inflates your context. So the more the more prompting you have to do for these different tools, the, like, kinda, you know, worse it gets in terms of, like, even Claude ability to use the tools. Like, you know, it it it's it's not obvious how this works, but it seems to me, like, as you increase the amount of information you throw at Claude at the beginning, it's gonna have more trouble, like, figuring out what to do and how to do it. So we've tried to, like, bunch all these concepts together. One really important tool, here is send message. You know? And this was part of the original agent team's implementation, but we've enhanced it now. So send message is the same tool whether Claude Code talking to one of its sub agents, one of the, you know, one of or one of the, Claude Code. on an on the same directory that as I demoed here. And in fact, the other thing that's interesting about send messages, we used to make this distinction between a running agent. and a stopped agent, and. we kinda got rid of that. Now Claude Code can message any agent. If it's running, it'll get injected as a little attachment to the current churn so it can be nice and responsive. If it stopped, it'll it'll restart it. So this was something where, like, you know, for a long time, people pointed out actually on Twitter I saw that, like, you'd see Claude Code, like, try to resume an agent that was already running. You know? It it was trying to keep track of the state of everything and whether it had to message it or resume it or whatever, and we just got rid of that. So some of the work here is trying to simplify things for Claude Code, so that, you know, it kinda makes sense, and and we can compress these concepts as much as we can. Yeah. That's great. Yeah. The the forever work is just, like, figuring out, like, yeah, how does Claude Code think and what what does it want? Like and and it seems like now you can sort of create your own agent topology. Right? Like, you can be like, hey. Like, I think agents are best if there's, like, a mailbox agent that, like, routes messages between the other agents and just prompt it, and they will, like, do that. Or you can prompt, yeah, like, a coordinator that, like, coordinates between them, or you can come up with, like, a telephone line or something. Right? Like, it's like extra like, people can build their own skills and, like, versions of, like, these agent implementations. Is that right? Absolutely. And I think if we've we've seen anything, it's that, like, these really differ. You know? So. I'll just describe some of the different things I do. Like, Yeah. one thing that's really common for me is I'm working with Claude Code, and I realized, oh, you know, there's another bug related here. It's, like, kinda too big for a sub agent to just do it one shot in a work tree. So what I'll do is I'll tell Claude Code, you know, like, using some of the automation I have here, like, hey, Claude Code. Can you open a new cmux window and, make a new work tree and give. the new Claude Code a prompt that will help it track down, the bug that we just found? You know, in that way like, I used to have do this by having Claude Code write an MD file in a temporary directory and then, you know, loading it loading it in. Now this is much more automated. It's it's very flexible. And I think by you know, one thing that's always true with with when building products is, like, you take this jumble of ideas that seem to kinda work together, and, like, your v one is often a bit of a hodgepodge. Yeah. You know? Like, you you have all these different things that go together, and it's not clear which ones are essential and which ones are just kinda things you like or you know what I mean? Or that was a ladder that got you to the really core insight. So I like to see a v two that, like, boils things back down into primitives. Like, list your peers, send a message, resume and send are the same thing. You know? So, ideally, I'd like to deliver all of the functionality and more, but, you know, using a smaller set of truly primitive things that then you can put together how you want. You know? Like, you should be able to use, multiClaude send message on the same computer whether or not you use tmux. You know, those things don't have to be bound together as an example. Yeah. This makes a lot of sense. And I I think, like, just this, like, overall idea of, like, how do you like, to find the right product, you really just need to, like, ship and iterate and talk to people. And then, like, I think probably what you see is that initially you do an implementation, and it's kind of, like, rough, like, agent teams, and we refine it. And then I think as it goes, like, Claude gets better at using it itself. You know what I mean? Like and then you're now it's like Claude is just, like, you know, creating the agent teams on the fly. You're not even thinking about that. Right? Like and and so I think when people are asking about, like, how do you understand features, I think the ultimate end state for us is, like, Claude is, like, you know, like, managing this for you. And there are some intermediate states in between, but, like, so much of, like, the added complexity, ideally, Claude Claude knows about too. So Yeah. So and I should add that so much much of it comes from dogfooding internally. Like, bef. before anyone outside ever sees a feature, we've subjected our internal users to it for. for a lot. And, you know, it's it's people sharing very specific feedback about things that go wrong that help us, like, tune things like the prompting, the triggering. You know? Like like, getting it so that Claude uses the right amount of sub agents is is tricky because people have different preferences, and, you know, it's it's not obvious that, like, oh, this saves tokens or it spends tokens. You know, it it kinda depends. The the frontier of that is, like, very squiggly. Yeah. That's right. Okay. Let's I feel like we can talk about we do an entire webinar just on agent teams, but let's let's, get to everything else we've shipped. So, okay. We've got agent teams. Yeah. Other things we wanna talk about. Let let's start with 1,000,000 context. So we we ship 1,000,000 context by default to every subscriber on Opus. And I think it's I'm curious how your experience has been. Like, has that changed how you, like, you know, how you prompt Claude or, yeah, how you work with it? Yeah. Interesting. I mean, you know, the opinions differ tremendously even on the Claude Code team. Right? Yes. So, like so, you know, I know some people who are like, I love compaction. I have one agent. It's been running since January, and it just backed you know I mean? It's like, everything's great. Yes. I personally am, like, kind of a neat nick when it comes to context. So Yeah. when I see a session compact, I'm like, oh, that's it. It's over. You know? So for me, 1,000,000 context has been great. And a big reason why I am so, in favor of sub agents is because they save context. You know, they prevent, like, all this extra context from accumulating. Yep. However, you know, there are plenty of times where, like, I'd much rather keep going than compact or start a new session. And at those times, you know, two hundred two hundred thousand tokens, even when you're using sub agents heavily, it's not enough to do some things. So I do rely on this feature. I do think there's something to this thing. Like, when the session gets really long, it's just unwieldy. If. it if it gets cold in cash, restarting it is gonna be expensive. And, you know, the performance does seem to go down. I think there are more cases than people realize where you'd be better off Chris everything you've learned as markdown documents or or in progress commit or something like that and starting over. You know, especially if you have the option of having Claudes to each other, you can have one Claude see another Claude and make sure it it pulls all of the context out. So I. think that will help. actually interesting. I I don't thought. What do. you do, You can roll your own compact that way. Right? Like, you can be like, oh, like, spawn talk to new Claude session. Here's the context. Right? Yeah. I I think Yeah. And then the old Claude can go back and be. like, oh, but I don't understand this. You know, that's that's something that's great. It's like, don't have to one shot the prompt for the new Claude. I I hadn't thought about that at all. I I think I was definitely very much, like, a big new session believer. I think before Million I was like, okay. Every time I'm making new session like, the the state is in the code base. Like, just read the git diff or something. Right? And then and then keep going. I think some people said when we launched MillionContext 4.6 that it felt like a different model. And I I think that is true to some extent where it just, like, it's better at these longer tasks. It can do things, just more reliably over, like, yeah, over a longer horizon. And I feel like I've been using clear much less, and I I've still been, like, pretty happy with the results. I think that, like, it's worth calling out as well. Like, when you're evaluating models across these different domains, like, I think it's probably true that there are some things that MillionContext is quite good at and some things it's bad at. You know? And so, like, we did a lot of evals on, needle in the haystack. Right? So, like, you know, this prompt I have here where you're like, oh, hey. Give me a a 100 PDF and find this, like, one or two, like, pieces of information, you can kind of only do this in million context or, like, it might be even better than, like, like, if especially if they depend on each other. Like, it might be kinda complex to do with sub agents. And I think that those evals are quite good. But then, yeah, maybe if, like, you're working that particular type of, like, coding environment or my my guess is that if you're changing your intent very often, inside of an like, a coding environment or inside of one session, maybe, like, that might confuse the model. So I think this is one of those things where you really have to get a feel for the texture of the model and, like, how it's working with you and, like, what your habits are and then decide. But I think 1,000,000 context gives you more options. Like, you can, you know, compact earlier. You can switch. Yeah. You can now, like, do handoffs and things like that. So, yeah. Yeah. Any more million context thoughts? Or no. Okay. Let's keep going. oh, okay. Okay. I I think, like, either I'm freezing slightly or you're freezing every now and then. So, okay. Memory. So we we shipped auto memory to Claude Code. Right? And so this means that Claude will now, you know, like, read and write memories. They're just files or markdown files, but I I think the way to think about it is that claude.md is your instructions for Claude, and then memory dot m d is Claude's, you know, memories for itself. Right? And so, like, we don't really think that you should have to manage that. Claude will do this. But, yeah, what what's the been your experience with memory so far? Yeah. This is really interesting. You know what? I. think it actually is a good example of the difference between, like, something that could. be good and something. that is really good. You know what I mean? Like, I totally agree. Yeah. Like, I shouldn't have to manage memory myself. And yet at the moment, I am tempted to. You know? And, like, I think I think you know, at times, memory has just absolutely astounded me. You know? Like, I see I remember something, and it's like, oh my god. That's so good. You know? Like, I'm so happy that I remember that. I didn't have to explain it. again, and it it was really useful. However, there have also been times where Claude remember something that, like, isn't right or, you know, like, it it sort of sends it down the wrong path or you know, I think we've probably if you've used memory, you've seen this thing where, like, you tried to implement it one way. Claude remembered something about that. And then, you know, like, you you you do it again, and you actually really want a. fresh start. And we can also make it really hard to test, like, a sub agent feature or something like that because, like, you don't really know if the sub agent is getting fresh context because it sees the memories. You. know? So it can make Claude's behavior more unpredictable, which, you know, that is that is really. tough. Like, Claude is already pretty unpredictable, Sure. and sort of adding these things that make it more unpredictable is is always a little scary to me. I'm glad we're testing it. I'm sure that at the end of the day, Claude will manage his memories, and this will be an awesome feature. I also am pretty sure that it will be much more complicated than we hope. You know, that it'll we'll have we'll have notions of, like, working memories and, like, solidified memories and, you know, ideas like dreaming. Like, all of these things will have to get worked out. And, you know, we need your help to test them, but I think right now, I think, you know, it it's probably net positive for me, but barely. And I think, you know, getting these features from, like, they kind of work or they can work well if you know how to use them to, like they just work has has proven to be a lot of a lot, you know, in it for all of them. Yeah. I know memory is one of those things that we have tried so much. Like, we've tried so many different iterations and, like, tweaking and tuning. And I think that, like, you know, Robert, I I I just saw him go through this entire process where he was really, like, very rigorous about, like, hey. I'm only shipping this if this is, like, net positive. And I think we, like, saw it in all the metrics where people are like, okay. Great. Like, this is this is helping. You know? And and and, I I think it's, yeah, that balance of, like, definitely, you have to, like Claude will get better and better at memory. Right? And so, like, right now, you have to build a little bit of an intuition for, like, should I ask Claude to write memory or edit its memory or something? Right? And this is, like, just a a little bit of, like, you know, your relationship with Claude Code. And then as Claude Code gets better, you it's know, like, yeah, perfect memory. You forget, you know, that you, like, even, like, ever had to tell it to write write a memory or something. Right? So, yeah, I I think, overall, I have been like, initially, I was, like, not sure how I like, I was like, oh, like, I feel like I need to read every memory file. And not not gonna lie. I don't read them now, and I've been mostly happy with it. So, yeah. Yeah. I think, for this one, I'm I'm curious to see what everyone else's thoughts are. And, I know people give us a bunch of feedback as well that we're iterating on. So Yeah. This is where we're, like, the tuning the prompting, you know, is, yeah. like, weeks and weeks and weeks of work and and. just getting it getting it to feel good, you know, and and and so many cases that you don't think of when you first implement it. Thoughts to do. But I. like that point that, like, these things seem to somehow move into the model. You know? Like like, they start as these product features and, like, over time just sort of become more and more natural till they go away. You know? Like, I feel like plan mode is kind of on its way there. You know? Like, there's a while where, like, if you didn't use plan. mode, you you'll, probably go away someday. yeah. Yeah. I I think you might have cut out, like, a little bit there, but, yeah, like, a plan mode. Now Claude Code has a enter plan tool. Right? So it can enter it itself, and and it will do it. And, yeah, eventually, it just, like, thinks exactly the right amount of time and, you know, does does everything right. So, okay. So that's memory. Computer view use versus Claude and Chrome. I I think this is about one of those things where computer use is another research preview that we rolled out. I I love this. We, like, acquired a computer use company, and then four weeks later, we have computer use in. It's like, you know, I I think really how we do things. But I I think this is one of those things where we're shipping a lot of primitives here. Right? We are shipping computer use. We have Claude and chrome. You can, of course, ask Claude Code to, like, create Apple scripts and things like that to control your computer. There's, like like, dispatch and things like that. So, yeah, I'm I'm curious, how you're thinking about the the, like, computer use primitives and browser use primitives we have so far. Yeah. Well, this this feature is so cool. And, you know, I think when you see Claude Code using the computer, you just know, like, Yep. okay. This is gonna happen. You know? Because there are just gonna be so many things that we never quite get an API or CLI for, and. this will be the way that, like, Claude Code uses a legacy system. I personally haven't had the need for it. You know, I I do so much work in the CLI. When I'm not doing that, I tend to use web. So I do use the Chrome integration for Claude. And it's funny. You know, when that first shift, I was like, that just seems so. dangerous. Like, I I I can't imagine ever using that. I I I don't trust Claude that much. Then I had this really complicated flight itinerary I had to book, and I was just getting so frustrated. I was like, alright. Screw it. You know, I installed the Claude, Yeah. Yeah. and I gave it, like, every permission I could. You know? It's just I went immediately from, like, I should be careful with this to, like, YOLO mode, which, you know, I think it's pretty bad. I think a lot of us are, like, not thinking about these permissions hard enough, Yeah. but I've seen, you know, in myself how it goes. But ever since I did that, now I use Claude and Chrome all the time, and it's really cool and really useful for you know, Claude Code can control Claude and Chrome or at least talk to it, and that that works out super well. So, I found it really cool. Actually, my favorite trick is to just install Playwright and then and have Claude write scripts so that it can kind of render a page. Like, that's often enough. So think some of these things are very cool, but they can be very high latency. And for me, getting caught in tighter feedback loops is always is always the goal. What what what are you doing? Yeah. Well so, I mean, Claude Chrome has a JavaScript tool now, which I actually have found to be quite good. It's like the equivalent of, you know, writing Playwright. Right? It can execute JavaScript on your thing. And so I I have been leaning towards that. I I think computers I I did set up a, Mac Mini. Like and so with channels in computers, I I think it's quite, like, they're just some things that only computers can do. Right? Like, if there's no Apple script or there's no, like, browser or something. Right? Like, I think I had, like, my WhatsApp messages, and there was, like, no like, I hadn't logged into WhatsApp in my browser, and so I was just, like, asking, and I was trying to do that. You know? And so, yeah, I I think that there is, it just ultimately is, like, the surface area of what you're trying to do. You know? Like, if you're doing these, yeah, local automation things, you're in front of your computer, probably like, computers is probably not going to be faster. But, yeah, you're touching grass in the park and you, you know, you got your Mac Mini. It's definitely going to enable some things that you just could never have done before with with Claude Code and just with Claude Code and Chrome. Right? So, yeah. But, yeah, definitely, again, one of these primitives, and we're just, you know, we're like, how how how much cogen is there? How much visual? Like like, I think there is a probably gonna be, like, evolutions there. So, yeah. So research preview, computer use. Okay. Claude Code Remote. Okay. So, we've shipped a lot of new things for Claude Code Remote. I think this is something that people have actually sort of been sleeping on a little bit because I think some of the initial limitations for Claude Code Code Remote were that, you couldn't get your environment easily set up. Right? So, like, your local environment and your remote environment were different, extremely different sometimes. There were some reliability stuff there too, And, we've, like, iterated on a lot of this now. So, like, you can set up in a, like, web environment. Claude Code Code Remote now has GitHub, integration, so it can, like, actually use GitHub CLI. It can do things like auto fix or PRs. Right? And, it also has yeah. You can teleport background jobs, things like that. And this is just like yeah. Like, I think a lot of what I was hoping for from CCR, you know. Yeah. And I'm curious, like, like, I I think you're, like, a local dev person right now. Is that right? Or Well, let's. see. Local dev on a on a home space. Let's see. There's a bunch of there's a bunch to to tease apart here. You know, I think, I guess, you know, starting with, like, why why has the CLI been so resurgent? You know? Like, why are we back in the terminal again? I think the answer is that, like, giving Claude Code access to the computer is really powerful. You know? Like, it it's easy to teach Claude Code to use the the computer, especially, like, a shell like Bash or Zeesh because it's, you know, it's it it can type it out and see the responses, in, basically, its native token token y language. Yeah. So I don't think there was anything in about the UI that was inherent there. It was about the locality. Yep. And, you know, everything about the web UI is better. Like, it you know, especially. It says. these multimodal things or, like, anytime you have to read one of Claude Code's plan in a terminal, I I mean, I would so much rather be looking at that on my web browser. Right? Yeah. And especially, like, commenting on it, giving in inline feedback. So I do think that GUI is the future. The thing about Claude Code Remote is just, like, it's very hard to have a container reliably there. You know? Just, like, just start there. Like like, starting it, stopping it, resuming it. Like, these things are totally nontrivial, especially at the scale where we operate. So, Yep. of course, the first version of this feature was pretty rough. We already established it. Like, we don't really know what these features need to do before we ship them. So you had to be very motivated, I think, to to try the first versions of Claude Code Remote. Increasingly, it's getting to the point where Claude Code Remote does just kind of work. And, you know, for the you have to develop a sense of, like, what's gonna work. There are a lot of things where, like, it's just not easier to have Claude write a bad PR, and then you kinda have to fight with it and pull it down locally anyway and test it. But there are some things where it's like, hey, Claude. This is an easy one. You should do this. I know I'll I'll I know a few people, do this sometimes, like, use Claude Code Remote as a task list. Just, you know, hey, Claude. We're gonna have to do this feature. Can you just do the research and do a little report? And then I can go paste that in when I'm ready to to take it on. So there's a lot of interesting workflows here, and it's one of these things where if you haven't visited it in a little while, you might be surprised at how much better it's gotten while your buck is turned. Yeah. Yeah. I I think that this is, like, definitely, like yeah. In the Claude Code repo, we've got all these verification things and, like, you know, we set up the environment really well. And and I have definitely been surprised. And one of my favorite features is a multi repo things because, like, this is kind of a pain locally where you're like, oh, like, I need to, like, clone two or three different repos and have them, like, coordinate. But, like, if I'm making a change in Claude Code and then also in the, like, docs repo or something, that's, like, very convenient, it can sort of, like, do, like, both of them. And and so that's one of the use cases where I've been pretty happy about. But I I do think that, like, yeah, if you haven't played with, like, Claude Code Remote in a while, this is, a good time to, like, try it out again. So, yeah, I think, let's I I I do have let's we we got one couple more things. I did want to do a quick interlude for by the way, because we didn't talk about it. And and and so. I meant. to mention. that. Yeah. I'm doing, my demo. Thanks. yeah, Sorry. just very quickly with what's. happening with by the way. It's a very popular feature. Yeah. Okay. So, one of the great researchers here, as just like a side quest, added this feature to Claude Code where you type slash BTW and then you give a prompt. Yep. And so the this was, like, kind of you know, it's a lot like fork, but without the join. So this. is great for a side question where you don't wanna derail the main the main Claude, but you do, you know, just wanna get one one follow-up, and it it shows up in a little panel. I think, I'm glad to see so many people using this feature, and it's the exact same semantics as fork. The difference is that fork is fork join. You know, Yep. Claude goes off and does a bunch of work, and then its final response gets relayed back to the main agent. By the way, it's just fork, with with no join. And. I think probably the destiny for that is for it to become a proper subagent steerable just like the other subagents that I showed. That's what I'd really like. So you type b t w, maybe you give a prompt, and then now you have a Claude that will whose responses will never be seen by the main agent, but you can go off and and keep talking to it. But if you haven't tried BTW, that's a preview of of some of these other fork ideas, and, I think, I think you'll love it. You know? It's it's really, it's really great. You know? There are so many ways to use Claude, and the way I hope you're using it is to just learn so much. Claude knows so much about computers, and I think, you know, you must be curious because you can easily just sort of fall asleep and just be like, well, Claude will do it for me, but it's not gonna be as good. So I hope you'll check that out and use it to learn. Yes. So let let's do another five minutes of, like, HTML. Like, we'll cover HTML and no Flickr, and then we'll we'll do some questions. But, yeah, speaking of learning, right, this is something that I think both the you and I and just the entire Claude Code team has been really, like, interested in is, like, asking Claude Code to make HTML artifacts because, yeah, like, there's so much that like, when you were spinning up agent teams, I'm like, dude, like, I don't even know what a Unix socket really is. You know what I mean? And so, like, I asked Claude Code, you know, to be like, okay. Make me an explainer, right, so that I can understand your code. And then, yeah, just made me, like, this beautiful HTML artifact. Fun fact, this entire deck is an HTML artifact made by Claude Code. You know? And so, like, I just asked it, like, hey. These are the questions that, like, the features we wanna talk about. Could you make something pretty? Like, here are some components you can use in a different folder, and and it just did it. So, yeah, I I think that I'm increasingly getting excited about just asking Claude Code, like, hey. Make me an HTML artifact, and, using that for exploration, for brainstorming, for learning, you know, for understanding. Yeah. What what what have you been using for? Sure. There's this mini trend, I think, of, like, people playing with this. And, you know, Yeah. you can take it a step further and make it interactive. So not only. does Claude generate HTML, but it serves it. I like to use BUN for this. So it serves it using a BUN hot module reload app. Then you can add interactive features. Like, make it so that anything Claude sends you in HTML, you can click on it and comment on it. And then Claude Code goes back and intercepts that. You know? So especially using MCP channels, like, there's so much you can build here now. And I do think that, like, part of the joy of this whole Claude Code thing has been the way that it lets you customize your tools and your work flow and keep them really sharp. Yep. And that frontier is continually moving away just because of good ideas, not even just be not even because of capabilities. And. asking Claude for ritual richer visual explanations, richer artifacts that you can interact with in new ways, this is one of those frontiers that we're pushing on. And, you know, the the fun thing about this is, like, you can mix these things of, like, oh, hey. Make it so that when I press this button, you know, you do a little dance. And by the way, this content, you know, needs to get fixed in this way. And I think when you find workflows where you can mix those two things really freely, that's when it just feels like you're flying. You know? You're you're sort of like, hey. Fix this thing in the content and never do it again. And, you know, Claude's like, gotcha, bro. No problem. Yeah. Yeah. I am yeah. No. I I think there's so much here to to, like, I think, figure out how we can, like, basically, like, learn more from pod, understand more from pod, like, interact more visually and and rich. And, yeah, I think we're gonna do more here. So, so quickly, Claude Code no flicker. This is a new experiment we are running. You have to opt into it with this end bar. You saw Adam do the demo earlier about this. But, basically, it means that Claude Code is, like, a virtual viewport, I think, is the technical term. I forgot. It's a new way of rendering on the terminal. This means we can do things like the the bottom bar will stay at the bottom as you scroll up, and you can use the mouse. You can click and like that. But there are also some challenges, like copy and paste is weird. Like, the copy and paste keyboard command doesn't work anymore. You need to, like, just select it and and it'll copy for you. But, yeah, I've been using it, and I've been, like, pretty happy with it. I'm excited about other, like, UI stuff it does. Like, and it sounds like you are as well. Is that right? You know, I actually love the old way. And I just you. know, it's also such a technical feat, you know, Yeah. to, like to to be able to render, like, true CLI style this, like, incredibly interactive and and rich experience. Yep. I just found that something about tmux kinda gets in the way, especially when the transcript gets long and this fixes it. And Yep. I you know, we we we've done a lot of work to try to make it better, but this is a blanket fix. It gives away a lot of things. You know? It it down to, like, not really you know, one thing I think I'd love to see us all do more is, like, jump in and out of Claude Code a little more. You know? Like like, you're working at the command line, and you're like, oh, I just wanna remember how this bash command works. I want it to feel really light that you can say, Claude, quote, how does that bash command work? You know? And hit enter, and it just like, the answer appears right in line. So I do think we lose something with this, but it's worth trying, especially if you're you're having trouble with your terminal, whatever your setup might be. It it it does fix a lot of things at the cost of some others. That makes sense. Yeah. Yeah. Trade offs for sure. Yeah. GUIs are GUIs are great as well. So, okay. So q and a, we've got a bunch of questions. I'm just going to to read them out, you know, and we can we can we can start. So the first one is, like, how are you handling PR reviews? Seems like that's become a bottleneck. We we launched a code review feature as well. But I know this is just something that we've been thinking a lot about, like like, your thoughts, Adam, especially, you know, having coming from, like, you know, Meta and Robinhood where I feel like you guys probably did PR's Pieroo or I'm sorry, CodeRiver quite differently. Right? Yeah. So, actually, I have a little story about this. You know, Yeah. when I first started working with Boris, Boris put up Boris put up these giant PRs, you know, with, like, four features, a bug fix. You know? And I was like I was like, what is going on here, Boris? You know, how do you expect. me to review this? Like, this is insane. He's like, just don't worry about it so much. You know, just, like, look at look at the types, look at the tests. You know? Like, does does the amount of code look right for what what it's changing? Does it match the PR description? And, Yeah. let's stamp it. And, you know, this was totally alien to me. I was like, hey. You know, arguing about the name of a variable or whatever is, like, a good thing to do. You know, this is how teams are forged. Yeah. But Boris, like, in a lot of things, he was just completely right about this. Like, you know, at at the limit, Claude Code is going to rewrite all of this code. It doesn't really matter if it's the perfect variable name or if you're handling the four loop exit condition exactly right. Like, that will come out in the wash. Definitely keep an eye on, like, changes to types, changes to comments. You know? Like, I think these things end up being really, really load bearing. When I see prompt changes or Claw dot md changes, I definitely. review those very carefully. Yes. But I also stamp a lot of PRs and just figure, you know, like, we wanna have a release process that catches these things so they can't if they're if they're problematic, they can't ship. The test catch it. The dog fooding catches it. You know, we we've got all these stages in place. So I think, you know, I think the thing that Claude Code does is it lowers the bar for really complex implementations. Things that were out of reach strategies that were out of reach for the typical team in the past, like property based testing, TLA plus verification, end to end smoke testing. Like, these things are expensive, Yeah. but Claude Code really lowers the price of them. So I think if you're not investing in these things to give yourself a little more peace of mind, like, that's probably the way forward. Yeah. I I agree. Like, I I think there's so much I think it's worth just getting, like, some of your best engineers and, like, we we've done this or to just be like, okay. Like, how are we verifying Claude codes work? How are we giving it, like, measures of, like, you know, you have things like latency and and memory usage and all of these different things. Right? So that, like, every it's really huge multiplier on everyone. Right? Like, you just feel, like, a lot more confident on it. And it's not like this is like, you're saying, like, something that you wouldn't want from it's not a result of, like, like, you would have wanted this in previous organizations as well. It's just that code was so expensive and it's so rare that you're like, how can you spend, you know, like, time doing, like, your testing suite? Like, we just need to ship the product. Right? But now, yeah, you can, like you can write 10 times as much, like, testing code as, like, as product code even. Right? And this can be, like, yeah, playwrights and, you know, like. like, you can have Claude Code record a video of your like, we're we're you know, like, doing the work. Right? And, this is, like, one of the like, yeah. We we we do this. Like and and so, this is one of the, like, I think, really important things just like, maybe we're moving it seems like what you're saying is, like, there are some details that in code review that were important that are less important now, like, what you named as variable or something. Right? Because previously, it was, like, sort of, like, you know, it's all, like, humans interacting with this code and, like, you really need to, like, gel together and, yeah, like, work well together. But, like, some details you don't need, some things you need to pay more attention to, like system prompts and and CLAUDE.md's and things like that. And then you can take this time that you're saving and invest it in the, like, verification infrastructure and things like that so that you could, like, move faster as as the models get better. Is is that right? Yeah. Yeah. And as everything speeds up, you know, like yeah. And I think, actually, I think one, like, one thing that we've kind of seen is, like, having an opinion that differs from Claude Code Code's opinion about code in particular is unproductive. You know? So, like, you can continually try to get Claude Code not to use JavaScript classes or something like that, but, like, you're better off just trying to stay on distribution with Claude Code and worrying about other things. You know? Interesting. And, like, if you can if you can have a lint rule or something that's like, you know, I hate seeing ternaries form this way or whatever. Sure. That's. fine. But but you you kinda don't wanna fight with Claude Code too much. It's it's it and part of the reason why I think Claude Code code is able to move so fast is because it's very on distribution for Claude Code. Has It been from the beginning. Right. Okay. That yeah. I haven't actually thought, like, in-depth about that, but that that makes a lot of sense. Code review. So we shipped code review and, you know, like, I I think that one of the things that, you know, to call out there is, like, it does use a lot of compute. You know? And, my model for this has been that increasingly, like, the job of an engineer is deciding what needs compute sometimes. Right? You're like, oh, like, is this a risky PR? Like, does it have a security vulnerability? Use the call out code code review. Like, please spend the, like, $15 or something. You know what I mean? Like, it'll save you so much time. But then, yeah, like, if this is, a stamp or something like that where you don't need code review, like, the the intense code review bot, maybe maybe you still have, like, a GitHub agent, action that, like, you know, imposes some style and things like that. But, yeah, does that map onto your like, do you have code review as well, or do you see that changing as as things progress? The new code review feature is amazing. You know, Yeah. it's it's like it's expensive, but in my mind, it's worth every penny because it just catches these incredibly subtle things. Now sometimes I feel like it catches subtle things that, you know, maybe don't need to be fixed for the first PR for an experimental feature. You know, it'll be like, hey. There's a race con condition here. And then they'll describe the race condition. You're like, okay. I think I can live with that. You know? Like, let's see if. people like the feature first. But the things that catches, you know, that, like, really have saved our bacon, you know, is it like, you know, it's hard to count, you. know, the amount that. all. that. And I guess the other thing that I like about code review is that there was an open question of, like, will Claude Code be able to do this productively? And I. just think this feature definitively answers that question. Now the cost may be too high for you. You may not like its exact output, but, like, the answer is yes. And I I also think that, you know, you asked this question about scaling. Like, the place where we're all headed is that Claude Code approved some PRs. Like. and I think if you're not working towards that, you know, like, there are compliance questions, all kinds of big questions about how that should work, but that's what we're gonna need to do to keep to keep accelerating. So I think Claude Code I think, code review is our first glimpse of, yeah, that's gonna be possible. Because code review's pretty good at saying, like, here's a here's a big blocker. This is just a nit. You know? And and I think it it gets it gets that right. Yeah. And, I mean, we just put out the the mythos blog post as well where we're talking about, like, how we're finding security vulnerabilities and, like, a really important open source software. And and so, yeah, definitely, you can just see the curve of, like, the how models are getting better. It's, like, they're getting better and better at this, this skill. But, yeah, we weren't maybe sure that if they would. Right? So, okay. So, yeah, just wanted to talk. Another one was it'd be helpful to have a map of capabilities of Claude access points, like features that are, like you know, we have different parity between CLI and desktop and web and things like that. I I think that's, honestly, a good call out. And I I'm not sure if if you have any thoughts on, like, you know, how the team is thinking about that, but, I do feel like it's a pain point. Yeah. This is one of the perils of of bottoms. up development. I I honestly couldn't map that out for you. I'm always like, do we have that one on that? You know? Or people ask a question on on Twitter, and I'll be like, yeah. We have that. And it's like, oh, no. Not in this interface. You know? I. the the main thing I'll say, I'll shout out, the the way we've been doing docs lately, which is, like, all of it's reactive. You know? There's no plan no grand plan where all these pieces fit together, but we do try to, like, keep the docs updated as they go. So, if you're not using the docs to try to help you stay on top of this, that that's a pretty good way. But, you know, I I think this will probably be a mess as long as the Frontier is moving as fast as it is. Because, Sure. like, just the the price you pay trying to coordinate these things is so much higher, than, you know, the value you get from being a little messy about it. Yeah. I mean, the docs are a really good point. We have an amazing docs person just, like, full time, only does docs, and, like, really knows everything in and out. So, like, definitely, like, don't read my tweets, read the docs. You know? Like, don't go straight to the source. Right? But, yeah, I I think, that's really good. And I think there's also worth calling out that, you know, different services are for different users. Right? Like, Claude Code Desktop is maybe a different user profile than TY and and the web and things like that. So, sometimes it's not intuitive. But, yeah, sometimes we launch a feature like, by the way, and you're like, we would love to have, by the way, in Claude Code desktop too. There's no reason not to. It's just like, you know, it'll it'll take a little bit of time. Right? So it's definitely something that we should, you know, think more about. Before we wrap up, I I think we have one more question on how can we understand our usage better. Right? And I I think this is something that I've been spending a lot of time on. I've been hopping on the calls with, with lots of users and trying to understand, like, yeah, how are people using their, you know, Claude Code. And one thing I realized is just, like, we've added so many sort of, like, layers of abstraction that Claude can can do, but it also means, like, you're a little bit more separated from, like, the work. Right? And so the your sub agents work, I think it's actually really, really good because right now, it's a bit hard to see, like, when you're spinning up, like, 10 sub agents or something. Like, what how many tokens are you using? Like, what are they doing? Right? So I think something I've seen is that, yeah. Like, people like, we've sort of got more more and more sub agent use can be a huge token, like, cost. Especially, I've seen these sort of open ended verification sub agents. I I think where it's like, okay. Build it, like, test it, and verify. And the verification is not based on, like, a, like, make the bit build pass or something. It's just like, hey. Like, test if it works, and if not, give feedback and keep going. You know? And I I think, like, that sort of work, I think, is, like, hard for Claude Code, you know, or, like like like, can burn a lot of tokens that way. So that that's something I'd look at. I just, like, would really, like, read all your skills to understand, you know, how things are going. Whenever you're using any proactive mode, I think that can be a place where, you know, like, you don't, you aren't always aware of how the tokens are being spent. Right? Yeah. Hook and detect a ton of context too. Like, you know, we see a lot of people with all kinds of misconfigured hooks that cause all kinds of problems. Yeah. And and sometimes it's hooks also, like, maybe you needed it before and you don't need it now. You know? Like, there's, Yeah. like, a a change there. So this is on us as well. Right? Like like, we ultimately it's like, we need to make sure that you understand how your usage works, and this is something that I'm very focused on right now. I'm trying to, trying to help out here. But, yeah, I I think just, like, paying attention, especially, I think, to your skilled sub agents and productivity is is very helpful. So, yeah. I think yeah. That makes sense. I think let's see. Did we have any more, like, last questions? We just hit 1PM, so I think we'll we'll probably have to call it. But this is really great. Thanks for thanks for joining, Adam. I'm always hear. This to hang. out. Yeah. And if you guys have feedback on this, like, please let us know. I think we're always a trade off of sort of, like, how much structured content do we prepare, how many demos do we prepare, how much q and a, you know, and we wish we could, you know, make time for everything, but there's, like, always trade off. So, definitely, like, yeah, let us know what you'd like to see more of or less of, and, we'll we'll see you again next month. So okay. Take. care.