Claude Code 2.0 MASSIVE Upgrade! (NEW UPDATE) transcript

Over the past couple of months, Anthropic has drastically improved Claude Code, turning it into an exceptional harness for agentic coding. And when you are to pair that harness with state-of-the-art models like Opus 5 and Fable 5, you can accomplish so much more. Today, I'm back to showcase some of the latest updates making Claude Code even better. From cross-session messaging, which lets Claude communicate with other AI agents, to the new default auto mode, improved mobile app development, and so much more.

So, let's just get straight into it. Starting off with the new cross-session messaging. This was something that was introduced in Claude Code version 2.1, and this is something that is a really useful update because your separate Claude Code session can now communicate with each other. So, instead of constantly jumping between sessions and re-explaining context, you can simply just tell Claude to pass the relevant information along.

For example, if I have one session working on my front end and another one working on my payments API, I could just say, "Tell the session working on the payments API that user.name is now user.display_name." Claude can discover your other session using list agents and communicate with them using send message. And importantly, it's not just sending your entire conversation history or your files. It writes a concise text message containing the information that the other agent actually needs.

What makes this even more interesting is that Claude can sometimes do this on its own. If something changes in one session that affects work happening in another, it can pass that information across automatically. You can also run {slash} list agents to see which session Claude can actually reach. Right now, this is something that is available on macOS and Linux with this new Claude Code update. If you want the best AI tools, workflows, and drops before everyone else, join my free newsletter with the link in the description below, which is completely free.

But, the main use case are multi-agent orchestration and keeping parallel coding sessions synchronized. You can have multiple sessions take on roles and coordinate around shared state, keep front-end and back-end or separate work trees aligned on APIs and code changes, or treat named sessions like reusable specialists, letting one session research something and another pull that information later without re-explaining everything.

Now, this is just a pretty cool example of what's happening with this cross-session management feature. You can see that two separate agents on two separate sessions are playing chess, and this is the coordination that you can see with this new feature being displayed. Next up, auto mode is now becoming the default permission mode for Claude code sessions on Pro, Max, and Team plans. That means Claude can automatically decide when to execute actions instead of constantly stopping you for asking for permission, which should make agent to coding workflows feel much more smoother, and I honestly love this feature when you're working on basic tasks.

You don't need to have Claude code pop up and ask you if you need to do something. And importantly, the classifier calls that are being used by auto mode no longer count towards your usage limits, which is definitely the biggest add-on from this update. You can still switch permission modes whenever you want, and if you already have a custom or organization-managed default, Claude won't override it automatically. So, that's just something that you want to keep in mind.

But, you may be wondering, why did they even switch to this? Well, the big reason is because Anthropic tested 1,053 paid developers and found that humans caught only 13.6% of dangerous commands, while auto mode blocked 89% of it. Human performance also got worse as sessions became longer, and with auto mode, it stayed relatively consistent and from a productive standpoint. And Anthropic says that teams using auto mode are shipping roughly 25% more PRs, largely because Claude can run uninterrupted for a much longer session, especially useful with long horizon models like Opus 5.

And Anthropic also heavily red teamed auto mode. And if for those who do not know what red teaming is, it's their internal platform for testing and researching new models. But they essentially red teamed auto mode with the polar research, which is bringing in its classifier miss rate on their adversarial test sets down from 12% to 7% and in a separate third-party prompt injection evaluation, the Claude 3 will 5, Opus 5, and Sonnet 5 running auto mode had zero successful attacks across 720 attempts.

And this is essentially why they are making this the main default setting within Claude code cuz it's essentially just going to get better outputs and it's just going to be much safer for every user. You can now set Claude code's output style to concise, which is basically a mode that tells Claude to lead with the result, cut out necessary explanations, and keep its response much shorter, which still gives you the full details whenever you ask for them.

And this is actually pretty important because these long explanations can quickly eat through your token usage, especially across long coding sessions. And you can enable it through the config file, and this is where you can obviously set this output style to concise inside the settings. Now, it is a pretty simple update, but it is going to get you the most out of Claude code while saving you a lot of tokens. You can also see why people are calling this the beginning of attention interface instead of only thinking about how we interact with an AI cuz we're now starting to see how you can control an agent and how it interacts with us, and how much it can say, what it can prioritize, and when it can actually deserve our attention for something as simple as a Claude Code mode that could genuinely make Claude Code feel dramatically better to use.

Next up, Claude Code has gotten a bunch of smaller but genuinely useful upgrades to Claude Code itself, as well as the CLI tool and the desktop app. First off is where Anthropic is extending the 50% increase to weekly Claude Code limits through August 31st, and they're hoping to make that increase permanent, which is going to be awesome. But, they did mention that the demand is still putting pressure on capacity. Performance is also getting better across the board.

Claude Code desktop now launches roughly two times faster than what it did a month ago after Anthropic fixed some background startup throttling and added some smaller performance optimization. The Claude Code CLI also is using around two times less GPU at P99 thanks to a change that stops Bun's garbage collector from kicking in during active runs and instead of waiting until the process is idle. And probably the coolest update is the new design skill, which is in research preview inside Claude Code.

You can now use /design as a command inside your Claude Code session or the desktop app of it to generate editable UI artboards, choose a design, tweak it virtually, and then have it so that Claude actually implements that design into your application. Next is a pretty small but genuine update. It's a useful quality of life update where Claude Code desktop can now auto continue after you hit your usage limit. Once you reach the cap, just simply enable the auto continue checkbox at the bottom, and once your usage resets, Claude will automatically pick that session back up where it stopped.

So, you don't have to come back later and manually restart it. Moving forward, we have a really nice update, which is the iOS simulator, which is directly going to be integrated inside the Claude Code desktop app. On macOS, Claude can now build and launch your iOS app, and the simulator opens right beside the conversation, like artifacts. You can watch Claude interact with the live app, see it tap through screens, verify changes, or take control on the simulator itself.

That makes the whole mobile development loop much tighter. You can build, test, debug, and iterate on iOS apps without constantly jumping between Claude code and a separate simulator window. Now, this is a feature that is in public beta for pro, max, and team plans, and it is something that requires Xcode with the iOS platform that is installed, plus the Claude code latest update. Recently, the Anthropic team has made it so that Claude code can now also have a dedicated screen reader mode, which is a great accessibility improvement, where you can just simply run the Claude command for the screen reader, which replaces the visual terminal interface with a plain linear text that works much better with screen readers like VoiceOver and with other platforms.

It is something that also adds clear labels like you, Claude, and tools, and turns menus into simplified numerical choices. And you can even play a bell when Claude needs your attention. So, if you rely on it regularly, you can make it permanent through Claude code settings, or you can just simply change the environment variables use this. For the last couple weeks, we all know that Claude code upgraded its effort levels for {slash} code review.

Before every review basically used the same strategy. Now, Claude changes how deeply it reviews your code based off the effort level that you choose. At low effort, it is something that does fast single pass review over a diff, which is going to make it cheap enough to run before basically every push. Medium looks at the change code in context, runs on multiple passes from different angles, and it verifies findings before showing them.

I goes further by using fresh contact sub agents as finders and verifiers, whereas with X high, it is something that looks for downstream issues outside the code that you directly changed. And they also integrated the ultra mode with the code review feature, which is probably the most aggressive mode. It spins up a fleet of reviewers and agents, and it can independently reproduce findings to cut down false positives.

Anthropic says that they actually run this on every PR internally. Now, the chart that you see here is also a pretty nice telling on their human labeled code review data set. Claude Opus 4.8, as well as a couple of other models that use the slash code review, goes from about 17% of bugs found at low effort reasoning to 25% at X high. While the leading competitor goes from roughly 8% to 12%, so you're basically choosing the trade-off between speed and cost versus deeper recall.

If you like this video and would love to support the channel, you can consider donating to my channel through the super thanks option below. Or you can consider joining our private Discord, where you can access multiple subscriptions to different AI tools for free on a monthly basis, plus daily AI news and exclusive content, plus a lot more. An update that not a lot of people actually know about is that you can have it so that Claude code artifacts can now directly use MCP connectors.

That means you can build an artifact like a dashboard, internal tool, or an interactive app that doesn't just display static information anymore. It can fetch live information through the MCP connector, and once it is connected with that service, it can take action on behalf of each individual viewer when they are requesting it. So, the MCP is essentially turning Claude's artifact into a much more of a functional personalized application, rather than just a generated interface.

This is something that's also available for all paid tiers. However, MCP access is actually disabled for publicly shared artifacts, so just keep that in mind. The CloudCode desktop app also has a built-in browser now, which is a pretty big step for agentic workflows cuz Cloud can open up documentation, design files, dashboards, or basically any website directly inside the desktop app. It can then read through pages, click around, and interact with them the same way it can with your local deployment environment.

So, instead of constantly copying information from the browser, you can just have CloudCode work with it directly within this inner environment, and it is something that can actually go get the context itself and use it while it actually works on other things. So, that is a really cool feature, but something that you should know importantly is that the browser is sandboxed and it is configurable. So, you can control things like whether browser sessions can actually log into things with your own passwords.

It is something that can also persistently log in based off of the settings that you preset within your CloudCode sessions. Now, this is more of a tip rather than an update, but this is something that I didn't even know about until I had actually clicked on the visual feedback feature, which is essentially where you click on the pencil and you can actually provide visual feedback directly to your CloudCode agent. This could be to provide different sorts of improvements to the UI.

In this case, you can highlight a certain section by drawing or even writing a text or even dropping in a shape to highlight a certain area, and you can specify the change or even specify where you want to move this card or do something else directly based off the visual discretion that you provided. That is a pretty nifty new feature, which is going to help with the development loop. Now, I'll be honest, guys, the reason why you haven't seen a lot of the CloudCode update videos over the past couple of months is because I just switched over to CodeX.

It's a lot easier. But ever since they had improved the overall usage of the harness itself, improving the rate limits, as well as giving us this new incredible model, the Opus 5, which had actually generated this uh game that you see on the screen, I just started using the model much more and it's just much easier to work with than other harnesses. Cloud Code is definitely one of the best, if not the best harness that's available.

So, you would want to make the most out of it if you have these paid tiers like the Pro Plan or the Max Plan. It's just something that can exceptionally help you in a generative coding and you can make the most out of the Opus 5, for example, if you are to pair it with something like Cloud Code. So, this is just my two cents. I'd highly recommend that you use Opus 5 with Cloud Code, no matter what, otherwise you won't be able to get the most out of the model, even the Sonic 5.

That's about it, guys, for today's video on the new Cloud Code updates. I'll leave all the links that I use in today's video in the description below, but make sure you go ahead and take a look at the World of AI 500 platform. This is or also a really good harness that will get you insights on which models to use and how to prompt better. Make sure you also go ahead and take a look at the second channel. Join us on the newsletter, as well as join the Discord.

Follow me on Twitter and lastly, make sure you guys subscribe, turn on notification bell, like this video and please take a look at our previous videos so that you can stay up-to-date with the latest AI news. But with that thought, guys, have an amazing day. Spread positivity and I'll see you guys fairly shortly. Peace out, folks.