What's new in Claude Code

1. SPSpeaker0:00

   [on hold music] [applause]

2. SPSpeaker0:16

   Good morning, London. Good morning, folks. How is it going?

3. SPSpeaker0:20

   [cheering]

4. SPSpeaker0:21

   Yeah, there we go. That's the London energy we're, we're looking here. So yes, uh, my name is Ralph. I am part of the technical staff at Anthropic, and in my role, I work with partners and customers all across Europe, helping them to build on and with Claude. And when they asked me to present what's new in Claude Code, I was already prepared to do, like, a seven-hour session straight with everything that the team has been releasing in the last couple of months because if you have been following the change log, that's quite a lot of things we, we have to cover. But luckily for you and for me, uh, the venue doesn't have another day extra for us to have an event, and we curated a se- subset of features that, uh, we released in the last few months. So let's take a look into the two main categories that we are going to cover today, starting with the developer experience because, yes, we know that you're using this tool more and more every day, and we want this to make, to make you a pleasant and convenient experience while you're using the tools that you love. And the second bit, we are talking about autonomy, meaning we are empowering Claude to do more without interrupt you when it doesn't need it, so the feature... f- functions and features that let Claude do the work on its own. And last, because again, the change log sometimes looks like more a newsfeed with the amount of features we, we are releasing, I'm going to point you to where you can find more information, more details, and keep yourself up to date with the latest and greatest. All right. So let's start with the first one because even though Claude can do a lot on its own, we care a lot about developer experience and having this tool that is convenient, nicer to use, and pleasant to interact every day is paramount for everyone at Anthropic. So start with the first functionality, which is remote control. Can I get a show of hands if any of you have used the remote control before? Yes, very small fraction with a big audience. Remote control is a very nice feature that lets you start a session on your computer and take this on the go from your mobile app with Claude or a- in any browser in any other device. Think about it. Dispatching a session and go run some errands, walking the dog, that's something [chuckles] I do a lot, or maybe even a bio break if you want. Let's stop doom scrolling and start, like, Claude coding whenever we want. And it's a really convenient way to dispatch a task and interact with Claude every time it needs to come back with a prompt. You receive notifications on your app whenever Claude needs an input, and also it's pretty easy to set up. As, and as you saw in the audience, not a lot of people are using it, so why don't we take a look in a demo of how simple it is to set up? So I'm gonna start with a demo around... So what I have here on the screen is, on the left-hand side, I have my session running on claude.ai. So this is on, on the, on the website. And also on the side here, I have my Claude session that I usually have. So what you can do with remote control is just start with a command. Actually, let me make it bigger for you. Remote control, and just press Enter. Now your session is active in the web and on your mobile phone, and you can start saying, "Hey, g-" uh, just the same prompt, "Give me ten jokes about the London weather," 'cause that's what's happening today. It's a joke. It's spring, and we have, like, this raining outside. And there we go. It's started in there. And if you see on the left-hand side, I already have a session in here, which is kind of with a weird name, but I can take over my session from on the go on the website. What I can do as well, and something that you also should be doing, is that you can rename your session for... If you don't pl- place any arguments here, Claude will make sense of what you're doing and give it a name. Like, if I press Enter, it will tell me, "Oh, this is a London rain jokes." And now you know from your mobile phone that that's what the session was, was about. And you can keep on online. So let's say, "Okay, another ten jokes about the Tube." And there we go. You can have the back and forth between your session on the go and your session on your computer. All your tooling, all your dev environment, everything that you are doing on your computer is accessible from the f- both the web UI but also from the Claude mobile app. Very neat feature, very nice functionality. Just start using today. And if I can give you a pro tip that we use it on our side, you can set up on your Claude settings do- dot JSON file for remote control to always be on. So every session is a remote control accessible because you never know when your puppy needs to go for a new walk or when nature calls, right? All right. Let's go back to the slides and go to the next feature. And yes, I was supposed to ask for a show of hands, but I'm seeing a lot of people smiling and nodding that, "Hey, this is something that I was expecting." So yes, we finally have it, a flickering free experience. And turns out that in our previous, uh, our previous approach for appending data and new information into the terminal was a bit tricky to get right because when you have, like, a big scroll back on your terminal, every time you want to append something new, depending on what is, it is that's something, it might cause the whole repainting of the terminal, and that's why you see the flickering. But now Claude Code has a new full-screen mode which actually virtualizes the whole scroll back of the terminal. So you have Claude not only inside your terminal, Claude Code is the terminal now whenever you want to, to interact. That means that scroll back that causes all the flickering is now virtualized. We are only rendering the parts of the screen that you are actually seeing it, meaning even on very long sessions, we can keep the memory usage super flat.By virtualizing that scroll back and virtualizing the, the, this experience inside Cloud Code full screen, you can also... we can also start playing with other types of functionalities, like for example, clickable elements inside a terminal. So now you can have an option to both expand, uh, s-text and also, uh, other types of buttons that are inside the terminal. But better than myself talking about functionalities inside a CLI, let me show you in a quick demo as well. So I have here my Cloud Code, as you usually have, and as you see, it started in the regular, uh, s- in regular mode, not, not a full screen one. But now you can come with a slash command TUI. That stands for terminal user interface, and you have just the option now to pass the parame-parameter full screen. So with this one, Claude will reload, and now you are in a flickering-free rendering experience. Now the full terminal is Cloud Code, and now you can... I can give just two examples. So let's say, uh, write me a file with 10 facts about Europe and another one with 10 facts about the UK. And while Claude is doing all of this, you're going to see that normally what would you have is, like, the part-portions of the file that is gonna be rendered, and that's... you can only see, like, a couple or a couple lines. But now because of the clickable options, uh, clickable experience that we have, you can already start expanding some of these options. So I can already click here, for example, and open the whole 10 facts. And also you can see here down the bottom, I have an option to go way down to the bottom of my, uh, my conversation. In fact, if I just put here, uh, another instruction, write me a paragraph about the UK. I'm very inspired by the UK today. And, uh, you're gonna see that this option here for going to the end, now it's updated with, "Hey, you have a new message," right at the bottom. So you can click it and go straight down. So with this experience, we're... the teams are guaranteeing a flickery free experience whenever you're working with Cloud Code. Uh, I invite you to try on, and just like the other, the other functionalities we discussed, you can enable this by default. So every new session comes with a full screen UI, and that's how we are mostly doing ourselves as well. All right. Up next, we do have a whole UI refresh on Cloud Code desktop, and for those of you that have used the, the Cloud Code desktop in the previous versions, like a couple months ago and thought, "Yeah, not for me, not something I would use. I, I rather use the CLI," I'm with you. A lot of us do prefer the CLI for most of the things, but I do invite you to come back and take a second look because the team have revamped completely the experience in, in the desktop app, and there's a lot of new functionalities that you can use. You saw a little peek in the first keynote, and we're gonna see a little bit more in a, in a, in, in a, in just a moment. But we are seeing that other users are preferring the desktop app for very specific tasks. Um, I don't think we are going to expect you to use this on your full-time job, like the, the eight hours of the day or maybe 12 hours of the day, depending how much you're building. But for some specific bits, the desktop actually provides a much better experience. But let me give you a walkthrough in a quick demo. So I have here my desktop app. Let me just make this a little bit bigger. And first thing that you're going to notice is that on the left-hand side you have an option with all your sessions, right? Uh, and from the sessions you can start grouping them in different categories. And what I like a lot is to group them by projects. This will group every session by what GitHub repo or what repo you are, you are working on. So if I just click over here, you're gonna see that I have three sessions on my scholarly draw demo. The other ones, I have another projects working on, and these are grouped, so you can, for the ones that are parallel coding, using a lot of sessions, you can really see, uh, which one pertains to each, each project. Uh, the other bits as well, let me take here some, for example, one of these sessions that I was running. Uh, you have the option on the top to take a look... If you have a, a conversation, you can take a look at the plan that spin up that conversation. For those of you that use it plan mode and want to find a step-by-step implementation steps of a functionality, for example, you have the option to take a look really easily from the desktop app into what the plan looked like for that feature. So for example, I was creating here a dark mode for this application. The nice thing about the, this UI is that you can start appending comments on it. So if you select any part of that plan that you didn't like, you can just leave a comment to it. This will be pushed back to Claude for making that fix. If we go even beyond that, so let's say for example, I have here one application that I was, I, I, I built and I have a lot of diffs on it, so that's another point as well. You can start seeing the diffs of the, of the files from that pull request. Let me close this file one. And you can start appending comments as well for your code. So you can say, "Okay, uh, Claude, explain me this line," for example. And what's going to happen is that Claude will take that line in particular, uh, and start, uh, and start as, uh, checking what is that, uh, what is that particular, particular line and giving you an explanation of that piece of code. So, oops, let me close this one. Uh, yes. What else do we have on the desktop app? Yes, as Boris also mentioned, you do have the GitHub integration quite neat on the bottom, so you can append fixes into it, and even going beyond check for your files of a particular, uh, of a particular development, development project that you are doing. So as I was saying, this is not only the diff, but I can take a look at the files themselves. So if I click over here, and I have the full files of that project. So yeah.Take a look, explore a lot more. Uh, like everything else we are doing, uh, with, uh, with Claude Code, we are very open for your feedback. So experiment, report bugs, report new features, and suggest what direction of travel we should be going. And yeah, with that, let's go back to the slides. And we can cover... Well, we covered a little bit on the development experience bit, and now I want to talk about the autonomy part because each of the functionalities that I'm going to cover next have one major root cause. And do you know when you started a long-s-running session or wh-what you expect was a long-running session on Claude Code, just to come back and see that Claude got stuck into maybe a prompt, uh, authorization request, or maybe he got the wrong building command, for example. Just something silly that you thought, "Oh, Claude, come on. You could do this without me. I don't-- You don't need my, my interjection on this one." So yes, uh, models are getting much more capable, as we saw it in the previous session, and Claude is getting a lot smarter into assess when does it need your input or not. So yeah. Everything that we are going to show in this part of the session is to show you that is, is, uh, is our time to get more confident and that Claude can assess when does it need to bother you for any of those, those points that I mentioned before. And the first one, with the new modes that we have in Claude Code, which is the auto mode. Auto mode is that, for that specific case when you are working, think of, uh, pu-put a prompt into Claude and think, "Hey, okay, this is gonna run for, for some time," and come back as that au- an authorization request. "Hey, can I read this file?" For example, "Do you authorize me access this folder?" So that can-- That was quite frustrating, and with auto mode, those days are gone. What auto mode is doing is first leveraging the high capability of the latest models as well to assess if a situation is destructive or not. Because what's auto mode does is that for every time where otherwise Claude we will stop and go to you and say, "Hey, can I get permission for this bit?" Auto mode check that with two-- with a classifier that checks two things. So the first thing is, is this action destructive? Am I going to regret this down the line? So yeah, that's the first check. The second check is pretty much, uh, does this look like a prompt injection? So yeah, this... Is this part of the, the, the situation looks like that, uh, is gonna inject something on my prompt. So if the checks passes, so none of this is true, Claude takes the action on your behalf, so it doesn't stop and doesn't... it continues running. If any of those checks fails, what happens is that Claude will try to find a workaround, maybe a scenario. There's another path that we don't have to go and ask for those, these particular permissions and don't have that destructive, uh, way that we are going. Uh, and if that case is not, uh, as well possible, then Claude will come to you and say, "Hey, I need your permission to execute this bit." So this gives Claude a lot more autonomy and don't bother you when you don't need to be. So for example, we value a lot your time, and Claude should earn your time of that, that as well. Uh, Boris was talking about asynchronous coding. That's exactly the type of thing that we, we can start leveraging it. Let Claude run and just bother you when you, when you actually need it to be, uh, consulted for. Yes. The next one is worktrees. And can I get a show of hands if you ever used worktrees on your day-to-day? Yes. Oh, that's, that's much better. Yes. Uh, still, I mean, I would love to have the wh- the whole room working, but work- worktrees are your friend whenever you want several Claudes to work on the same project, same repo, and you don't want Claude, different Claudes to step on each other's feet. The idea here is that a worktree is a copy of your project in a segregated sub-directory that is only accessed by one Claude session. Actually, worktrees existed way before Claude Code, and they are supported by Claude Code since the early beginnings. But to be honest, it was quite tricky to get it right because you have to manage all these different directories. Worktrees are kind of ephemeral if you start working on, on something and then stop it. So managing all of that yourself was something that was quite tricky. But now it's super simple because Claude Code now has, uh, this native support for worktrees. Whenever you want to start working in a new feature that needs to be parallelized amongst several sessions, you can just start your session using Claude dash dash worktree, and Claude will copy that repo to a segregated directory and work on it independently. You can have multiple Claudes working on the same project in different features without them checking o- and editing the same files. If you don't-- If you're already in a session, for example, and don't want to start a new Claude Code session, you can just ask Claude to create a worktree for you. You will also see that in other UIs that we're going to present, uh, the desktop app, for example, and the other one that we're gonna show in a moment, worktrees are native to it. So if you start a new session in the desktop app, for example, there is a flag already saying, "Hey, this session is going to start a new worktree," so you don't have to manage yourself. So yeah, worktrees, use it. They're a friend. And yeah, if you're a para-parallel Claude Code user, like a lot of us, this is, uh, there's no other way to, to get it better. Yes. Okay. Auto-memory. That's another nice feature that is giving Claude a lot of autonomy as well. How many of you think that, uh, have, have thought that every time you started a new session on Claude, it looks like you are working on a blank slate? You need to tell Claude the same things a lot of times. And even though you have, like, your Claude MD file and some instructions to get it right, there are some little things that it's, it's never there or s- most of the times are never there. Auto-memory fix that. 'Cause auto-memory is something that Claude does automatically, so it will check for some notes, uh, and, and some, some things as your session progresses, and take notes in the back. So think about it, your coding style, some architectural choices, your debugging insights, for example. The way that only you know how, how, how it works. Claude will assess that and save this in a memory.md file. It's not going to update that file in every session. It's really have the discerning capability to understand, "Oh, this is something I need to remember for the, for the future."And it will load for every other new session together with your Claude.md file. And you might be asking, "Yeah, but we have Claude.md files, we have memory.md files. What are... Why, why do we have these two, and what are the differences?" I think an easier way to understand, imagine that the Claude.md is your- is the onboarding document you give to Claude on the first day of work, and the memory.md file is the notes that Claude is taking while you're actually doing the work. So one is a manual, the other one is a note-taking. Well, the other thing as well is that the memory.md file is a very short file. It's not like a big one. It's mainly an index to other files with the specific memories itself. So it's not bloating your context. Actually, we are explo- we are leveraging progressive discovery because every time Claude needs, for example, to get access to your debugging memory, you get access first to the index, find the file for the debugging memory itself, and then it loads in your context. So you don't have, like, this bloated context, uh, in, in, i-in, in every session. Other thing as well, all the sessions that are part of your project, all the work trees in that project as well leverage the same memory. And, uh, and also the memory files do not leave your computer. It's not something that's pushed to your GitHub. Uh, it's not something that lives in the cloud. So it sits on your machine. Last bit, you can definitely audit and check for those memory files by just typing /memory, and you can see the whole directory that Claude is managing for you. All good. All right. The next one, and this one happens with me a lot. I happen to work with partners and, and customers all across Europe, as I mentioned, and in a lot of these conversations, people ask me, "Hey, how is the Anthropic developer team using Claude in their day-to-day?" And of course, there's noth- not I can... n-not a lot I can talk about, but this one I can definitely do because this is a practice the teams have incorporated, tested, saw it was working, and bake it into the product. So code review is an automated review process for every PR that you, you create, and it's a process that is both multi-agent and multi-phased. So imagine that every time you create a PR, Claude will spin up a team of different agents to assess different parts of that PR. And, uh, these agents are checking for different stuff. Some are checking for errors, some are, some are checking for bugs, some are checking for vulnerabilities, logical errors, for example. And based on all of those findings, that's the first phase. There's a second phase which will actually assess all the findings from those agents against your actual code. So we can actually make sure that those findings are actually real. In that sense, we can speed up the findings of those errors much quicker. So you will find errors or logical, uh, lo-logical problems that will take hours for you to discover now in a matter of minutes. You can start, uh, the, the code review is already native on the GitHub app if you have installed Claude on your GitHub repo. So you can automate this for every new pull request that your repo, uh, receives. And you can also manually, uh, start a new pull, uh, a new review session using the command Ultra Review. It's not a Super Review, not a Great Review, it's a Ultra Review. /ultra-review from your Claude Code, uh, terminal, and you s- you're gonna see... you, you're gonna start implementing PR reviews the same way that Anthropic teams are doing. Yeah. Now for one that is really exciting, and we talked a little bit about it in the, in the keynotes early this morning, routines. This is sh- was shipped, uh, earlier this month, I think. Uh, it's still in a research preview. And this helps you create sessions in Claude Code without you have to manually triggering them yourself. So think about this, a schedule, a workflow that can be triggered based on an API call from another system, for example, or based on a timer, or based on a webhook. So you can start imagining all the capabilities that you can, uh, incorporate on your, on your workflows. This already works natively with GitHub. You're gonna see this in example. But as I said, you can just plug an API call to it. Think about every time your e-commerce makes a sale, you want to start up a, a, a routine with Claude, that's just an API call away. And the capabilities of the routines, the routines themselves are the same ones as an actual chat window, as a- an actually Claude Code session that you're having. So if you want Claude to make cURL requests for other systems, to check for some specific connectors and so on, it's, uh, it's what we are going to be able to do it. And, uh, yeah, why don't we see a quick demo of routines and some of its capabilities as well. So I'm back here on my Claude Code desktop, and as you can see here on the top, we have the routines button. So in this one, I have already prepared two of them. But, uh, as you can see, you can... I... Well, let's check, take a look at the first one. So for example, in this one, I'm checking a repository that's not my repository, it's just one, one, one pub- public repo out there. And I created one routine that runs both at 8:00 a.m. and 4:00 p.m. every day, checking for new commits on this master repo and also for new issues on it. If I scroll down, as you can see here, I also can, like, incorporate some cURL requests, right? You can do everyth- uh, you can ask Claude to do anything that Claude would do in a regular sessions. And down here, you can see we have it run every day. So this one already run today. And I have here the, both the new issues request and no new commits for this repo in particular. And you have all the login of every t- every day it was running without my intervention. My computer doesn't even need to be on because these are running in the cloud. So let's create one just to see how easy it is. So I'm gonna click on new routines, and here on this side, click on a new one, and you can create, uh, on a remote environment. That's what we are going to do. And I'm gonna say here, let's say issue, issue triage. AndLet's say assess the issue and give me your review. Do sound like a medieval knight, 'cause that's always more funnier li- that way. So we're gonna select a repo, a remote repo, and I'm gonna put here on my own scholarly draw fork, and my GitHub event, in this case, you can pick one that's gonna be an issue... Oops, sorry. An issue open. Not pull request, sorry. The issue open. That one. So every time an issue is open, Claude will execute this routine. You can also add some connectors to it. In this case, it's not applicable, so I just click on Create, and this one we have it connected to my repo, and so on. So what I can do is access that repo itself, which is this one, and I can create a new issue. Let's create here a new issue, and I say, "Hey, I found a backdoor on this app. There's a backdoor. Please fix it." I mean, there's no backdoor, but issues are random, so let's just create one of these. And what happens is that this will trigger a web hook, we will trigger my routine, and that will start. This take a couple of seconds, but while this is, is running, I can show you the other one that I created before. So I have here my posh MP review, which runs every time somebody creates a pull request on that same repo, and this one is reviewing this PR and comments with findings in the PR itself. So the one that I created is not interacting with, with the issue, it's just, like, providing me the report. This one is coming back to the issue and creating one review right there. So I'm clicking here, and I'm seeing all the first one, and I have this PR that was open. So I created a first PR yesterday, and Claude came and did a review based it on that routine. So it's not... Nobody has to tag anything, nobody has to, to make the integrations. Just the web hook from GitHub trigger a Claude session, and that started, uh, the whole PR sounding like a very posh MP. So that should give us time to come back to the one that I created with the issue. Let's say this one. Yeah, we already done this one in here, and this one is my medieval knight checking for the issue. So yeah, that's the kind of workflows that can start already building it. Think about a lot of the web hooks from GitHub, any API calls from your external systems, all of this running in the cloud without your computer being open. That's boils a lot to what Boris was, was mentioning that, hey, now you are asynchronous doing your work. Lets the Claudes do your work for you, and just come back and assess what is going on in there. All right. Uh, yeah, that being said, we can go really quick talk about Agents View. And you saw a little bit of this this morning. Uh, with Agents Views, you have one view for your s- all your Claude Code sessions. So if you are, like, an addicted to Claude Code parallel sessions like I am, you have that window with, uh, with that monitor with many terminals opens at once. Term-- uh, the Agent View lets you manage that in a much more seamless and easy way. You can see all your sessions based on what's their current status, which one are waiting for a prompt, which one are completed, which one needs your input, for example. And better than myself showing you a slide, let's go to the last demo of today so I can show you how does that work. So A-Agent View is really easy to get started. So if we just start our session with Claude Agents like this, uh, you're gonna see the first Agent Views i-is something that makes your sessions run in the background. 'Cause as you can see here, I already have two sessions with Claude Code open in there. And if I go a little bit smaller, you can see that I can start dispatching. So, uh, what is my schedule for today? Uh, give me a Spotify playlist for Code with Claude's London, and let's plan a new dark mode for this app. So as you can see, all of this is running in parallel and is actually running in the background, so I can leave Claude Code as I just did right now and come back. Everything's keep on running, so you don't have to be attached to the terminal anymore. You can go inside any of the sessions if you press Enter, and you can see here already have a Spotify playlist for London. You can come back. Uh, you can pick on some of this. So for example, calendar is there, so I can... If you just press Space, you can send a prompt to that session without delve- delving into the session itself. So, uh, how many appointments for today? And yeah, they are... Claude will-- If there's a session that is waiting for you to take action, for example, there's a group, uh, that you can see the grouping already from working to complete. In this case, for example, now session needs an input, and you can go over, over there, go into it, or, uh, just peek only to it as I, as I said before. So yeah. Agent Views is released in public preview, and I welcome you to try it if you are one person that uses several Claude Code sessions at once. So I could be talking a lot more. Actually, I'm already over time. And we could talk a lot more about other, other features in regards to Claude Code, but one thing that I have to mention to you is that none of this will be possible without first you leveraging those tools and using this every day, but also providing all the feedbacks, all the features requests, and all the bug reporting that you have been doing. So please keep on doing that. And for particularly the ones that are managing teams or enterprise, uh, deployments of Claude Code, we released several features that were made just for you. Better Windows support, wizards for cloud deployments with, uh, both, uh, Google Cloud Platform and AWS. We have native binaries, so you can leverage the installation straight from your, your, your govern- govern- governance pipeline. Uh, and yes, the last bit, just to be super crisp, is that there's a lot of, of other things and a lot of more that you c- we can talk about Claude Code. Follow the dev, dev team on the socials. We have the Claude devs on Twitter. We do have a what's news, uh, section inside the documentation that just sums up what needs to be assessed and what's, what's important for you to that, and of course, the changelog is there. Um, subscribe to our ne- our newsletter, and one thing that we should never stop doing, especially in this day and age, it's never stop learning. 'Cause there's a lot of things that we, we, we need to be aware of. So keep learning, keep building, keep thinking. Thank you so much. You guys were amazing. That was fantastic. [applause] [upbeat music]

Episode duration: 32:03