Antigravity 2.0 Just Changed AI Coding Forever 🤯

Added: 2026-05-23

[00:00:02][music] >> Hi. Welcome to another video. So, if you remember, there's a product from Google that's called Antigravity.

[00:00:11]I know that it's a shocker.

[00:00:13]Well, it hasn't been getting any updates. I mean, at least it seems so, but that is not exactly right. I can confirm that probably in Google I/O, you will see an announcement that will tell you that Antigravity and Jewels is now merged, and you can now make online running agents via Antigravity itself.

[00:00:30]I can say it actually confidently, because if you see then Google AI Studio is now powered by Antigravity, and so is some other stuff in Google Cloud.

[00:00:37]So, yes, they are probably not giving as much life into it, but it is not bad either, and some of the smaller updates that I'm seeing is actually really cool, and it is actually quite a bit changed from the last time that I severed it.

[00:00:49]So, let's have a look. The first thing is that it is just way more polished.

[00:00:53]Many of the nasty little bugs has been very much crushed. So, it is a really pleasant experience to use.

[00:00:59]The models and stuff is pretty on point as well.

[00:01:01]The agent contraption also works really well.

[00:01:04]Similarly, the settings and rate limits are not really transparent. So, if you go to the settings panel, then there are some really good new settings as well, which was really less the last time.

[00:01:14]To start on the left, you have the agent settings, where you can set up things like strict mode. When enabled, it enforces settings that prevent the agent from autonomously running targeted exploits, and requires human review for all agent actions.

[00:01:26]You can also set up review policy, terminal command, auto execution, terminal sandbox, shell integration, and some more permission stuff as well.

[00:01:35]Then you also get the notification settings. It allows you to set up the notification setting for when the work gets done, and so on. After that, one of the best features is the models feature.

[00:01:44]It now allows you to see the rate limits transparently about it. You can see rate limits for each model here, and you can also add AI credits to the usage over the limits if you want. Each model has their own rate limits, which refresh accordingly in about 4 hours, which is really cool to see.

[00:01:58]It is also available on the free tier with all these models and it being transparent for people to plan accordingly is really good move. Then we have the customizations tab as well.

[00:02:07]The customizations has option to set up skills and MCP servers. The Google related MCP servers are also directly available here if you want to use them.

[00:02:16]So it's kind of easy to set up. For skills, you can either add a whole directory from where it can automatically discover skills or so on.

[00:02:24]You can easily add maybe the Claude skills to make sure that once you add it in Claude or open code, then it automatically adds it in Anti-gravity as well. After that, we also have the browser option to set up the browser control options.

[00:02:36]For an example, you can enable the browser tools. When enabled, agent can use browser tools to open URLs, read web pages and interact with browser content.

[00:02:45]This allows the agent access to important and often critical knowledge and methods of validation.

[00:02:50]You can also set up the browser JavaScript execution policy. In disabled, agent will never run JavaScript code in the browser. In request review, agent will always stop to ask for permission to run JavaScript code in the browser. On always proceed, the agent will not stop to ask for permission to run JavaScript in the browser.

[00:03:08]This provides the agent with maximum autonomy to perform complex actions and validation in the browser, but also has the highest exposure to security exploits. You can also set up browser URL allow list, which allows you to control which URLs the browser can access. Add domains or full URLs to the allow list.

[00:03:25]You can also set up the tab and editor option as well if you want to do that.

[00:03:29]So that is the settings panel. Another thing that has changed quite a bit is the agent manager. The agent manager opens up and you can see that it is a bit changed than before for sure. You can see the models here to use as well and just like Codex, you can easily change the project folder from the simple option here. There's also the option of environment. It currently only shows the local option, but you can probably also have the remote option as well when it becomes available because it just shows the local option for now and the drop down doesn't make any sense.

[00:03:57]Another thing is that if you have a Linux machine, then it now allows you to use custom providers and models as well.

[00:04:04]So, you can potentially configure things like a llama and stuff. You also have the conversation history as well where you can see all your workspaces and conversations as well.

[00:04:13]There's also a new toggle changes pane where you can see the exact changes made and everything which is kind of cool for sure. As you can see the whole changes done in a session or conversation at a glance.

[00:04:23]You can also hit the review option which allows you to select a thing in your walk through and then comment on that which will allow your agent to fix that or change that accordingly.

[00:04:32]Then there's the option to rename the conversation as well to allow you to make it good looking and so on.

[00:04:36]Anti-gravity is actually really good now. The limits are good in the free tiers as well and it is really awesome.

[00:04:42]They have worked on it to make it just a more seamless experience which I appreciate a lot. So, yeah, it is actually really cool. There are also the rules and workflows option that you can open from the below anti-gravity settings option as well.

[00:04:53]Another thing that is not so big but still useful is that the docs are now much more fleshed out as well and if you actually go through the change log, then you can see why the app feels more alive than it looks from outside.

[00:05:05]The latest build is FOM April 16th and that one is mostly bug fixes. They fixed MCP servers not loading and they fixed workspace specific settings not being accessible. That is not flashy but for an agent IDE, this is core plumbing. If MCP is broken, then your external tools and custom servers are basically useless. Before that, April 7th added the unified agent permission system. I really like this direction because with coding agents, permission control is the difference between a fun demo and something you can use on a real project.

[00:05:34]You want the agent to edit, run checks and use the browser but you also want clear boundaries around what it can do by itself. Way then one version added Linux support for sandboxing, improved MCP authentication, simplified the chat UI, added support for reading rules.

[00:05:49]That is a very developer focused update.

[00:05:51]The rules and workflow side also got attention. The February updates are also worth mentioning because a lot of the settings I showed earlier are from there. They added the model screen in settings, which gives more visibility into quota usage. It also added the terminal integration setting, artifact downloads from chat, up and down arrow navigation for input history, and better UI responsiveness when creating conversations, sending messages, and reverting changes.

[00:06:15]That same update fixed one issue that honestly sounds scary. Reverting could occasionally delete files edited by the agent. So, when I say the app feels safer now, that is not just because there is a strict mode toggle.

[00:06:27]They are also fixing the basic trust problems around the agent editing your actual files. And speaking of safety version, they added terminal sandboxing for macOS users, while the March update brought sandboxing to Linux as well.

[00:06:39]Secure mode was renamed to strict mode in February, and the original secure mode update from December required human review for all agent actions and prevented the agent from autonomously running targeted exploits. So, the story here is pretty clear. They launched the agent first idea, and now they are building the guardrails and quality of life features around it. Permissions, sandboxing, MCP reliability, rules, workflows, artifacts, model quota visibility, long conversation performance, browser fixes, all of that is being improved piece by piece. I still think Google needs to explain the limits and credits better, and I still think the product has some preview energy in places. But right now, based on the change log, it does not look abandoned to me. It looks like a product that is being quietly hardened in the background. So, if you tried antigravity only when it first launched, I think it is worth opening again. Not because everything is perfect, but because the boring parts that make an agent tool usable for real work are clearly getting better. And I think that is probably the best way to describe this update. Not a huge relaunch, not a flashy new thing, but a lot more serious than it looked from outside. It is not yet at the level of Verdant.

[00:07:48]Verdant has been evolving a lot and making new features like the manager and everything, which is actually really innovative in my eyes.

[00:07:56]I think that Anti-Gravity is still trying to catch up and they have caught up a lot, but they should innovate more.

[00:08:02]They are still using the same old Windsurf things as well. So, a proper refresh is something that I'd like if it ever happens for sure.