Human
One IDE to bind them all
As with so much of its generative AI offerings, Google has struggled to offer a compelling alternative to tools such as Cursor, ClaudeCode or OpenAI's Codex. AI Studio, released in the middle of last year and which we reviewed at the time, showed some interesting potential, but the launch of AntiGravity demonstrates that - some important caveats aside - Google has not only caught up with its competitors but surpassed them in many ways.
Antigravity is an Integrated Development Environment (IDE), and indeed that is the way that I have been primarily using it for the last month - although it is perhaps better thought of as an agent manager. Certainly this seems to be the future for AntiGravity, allowing the app much greater control of testing and iteration, but in practice this was also one of the features that I found myself intervening in much more to check results (which does, rather, defeat the purpose of an agent...)
As an IDE, however, it is largely excellent, and indeed has become my tool of choice over the past month. It is based on VS (or, more accurately a fork of VS), and the standard view shows an explorer window on the left showing files or repositories, the coding window that dominates the view, and the agent which can be toggled on or off on the right. You can also call up a terminal view which is useful for connecting to remote files via SSH. Three of these windows (explorer, code and terminal) are very familiar and bring the standard tools that you would expect from a modern IDE, including excellent context aware suggestions and implementation of programming languages. The agent window is where, as with ClaudeCode and Cursor, everything has changed for developers.
From the drop down list of agents, you can select a number of LLMs to help you with your project, including Claude Sonnet and Opus, GPT-OSS and various flavours of Gemini. Many of the criticisms online about Antigravity tend to stem from the limited number of uses for various APIs; because I have a Google One account and tend to do most of my programming in the app with Gemini Pro - without encountering limits so far - this has not been a problem for me. But it remains the case that to get the most out of Antigravity it helps to have a paying Gemini subscription.
Lorem Ipsum has been the industry's standard dummy text ever since the 1500s.
While it easy to vibe code with Antigravity, to get the most out of it you should take full advantage of its ability to work in collaboration, clearly listing all the operations it will undertake.
Next level vibe coding - but with flaws
Because of its interaction with agents, Antigravity is gaining a lot of attention for a superb vibe coder environment. Certainly I've been using it to develop web-based apps and - 90% of the time - it is excellent, encouraging me to push my own skills. However, as with all vibe coding, there are some serious limitations and, to be honest, things will always work better if you have some core understanding of the language you are working with. This is not least because every once in a while the LLM will become fixated on some process or fail to spot an obvious mistake which can be quickly cleared up in the coding window. As with all AI, for all its brilliance Gemini sometimes cannot see the blindingly obvious.
If you are using it as a standard coding IDE, it includes many of the best features of VS-based editors with added improvements to make the whole experience as frictionless as possible. My one concern is that this is apparently a fork of VS: while that is great in terms of Google adding incredibly improvements in terms of its ability to integrate with agents, it brings with it a slight worry: Google can be brilliant in terms of tech, but it is also a company where the vast majority of its products fail due to neglect. I sincerely hope that Antigravity doesn't fall into the Google graveyard, but there is no real way to be entirely sure this won't happen.
Which brings us to Antigravity's incredible feature: rather than AI being an assistant, as in Github Copilot, it is an autonomous agent that can plan and execute instructions - and why it is being adopted by so many users looking to improve their vibe coding. When you enter plain language instructions, the agent interprets these and creates an implementation plan which you can then modify or agree, leaving it to write elaborate and extensive code. Antigravity's code assistance goes beyond autocomplete and linting. It builds a semantic understanding of the project, offering suggestions that account for architectural intent rather than just syntax. For example, it can warn when a new function subtly violates an established design pattern or suggest refactors that align better with the project's historical evolution.
It is not, however, without its weaknesses. The learning curve can be steep, particularly for developers accustomed to traditional IDE paradigms. The adaptive interface, while elegant, can initially feel disorienting, as familiar panels appear and disappear based on context. Additionally, some developers may be uncomfortable with the level of inference the IDE makes about their intent, preferring more explicit control.
To be fair, Antigravity regularly invokes your interaction, although it can be set to proceed much more autonomously - something that should be used with caution. Despite this, and the fact that its agentic control of your browser to test and implement features (indicated by a blue glow around the main broswer window) sometimes falls down, this really feels like the future of coding design. While it has not yet made the complete step from assistant to fully functioning agent, this probably feels closer than any other product out there although ClaudeCode remains our editor of choice.