Google Antigravity IDE
andre
Building an AI IDE is no longer a moonshot. It is merely a matter of wrapping a chat interface around the open-source VS Code editor. Think of it as two car manufacturers sharing the same chassis and mechanical foundations, rolling out slightly different but obviously competing models. In this case, the shared platform is Microsoft’s free-for-all code editor. Cursor recognised this reality early. They took something entirely open and free, slapped a few chat windows on top and plugged some AI pipelines into the back end, starting to charge twenty dollars a month. Then two hundred. It is a rather audacious business model, and they have been handsomely rewarded for their nerve with a towering pile of investment capital from Elon Musk.
One might object that the underlying concept is hardly complicated. The editor already exists. The AI models already exist. Stitching the two together with a conversational interface should not represent a monumental engineering challenge. One expects that a company of Google’s calibre could imitate the formula in a long weekend and ship something superior by Monday. And yet. Antigravity IDE stumbles on a number of very basic points, which suggests the task is somehow harder than it looks.
Consider the single-chat restriction. You can have exactly one active conversation at any moment. There is no way to open a second chat within the same editor window. You can, technically, launch a separate editor window and start a fresh chat there, but that is a deeply inelegant workaround for what should be a trivial feature. How difficult would it be to permit two or more simultaneous conversations?
Someone will inevitably protest that nobody actually needs to talk to two AI agents at once. This person has clearly never written production software. I do it constantly. In one conversation, I might be building a new feature or debugging a stubborn regression. In the other, a production incident has just materialised and demands immediate investigation. I cannot dedicate my limited human attention to both simultaneously, but that is not the point. Coding with an AI agent is an interactive process. You issue an instruction. The agent begins working. And then you sit there watching code diffs scroll past the screen, waiting for it to finish. That waiting period is tedious. While one agent grinds through a prolonged task, you could be doing something genuinely productive in a second thread instead of drifting into internet browsing. I could be checking TikTok, but I do not even have it installed on my phone, so the alternatives are grim.
Equally frustrating is the opacity of the context window. There is no meaningful indication of how much conversational memory has been consumed. Now, most of the time, developers do not stare at this metric obsessively. But occasionally it serves as a useful nudge: finish the current task and migrate to a fresh context before the model starts degrading. This limitation will persist for a while under current AI architectures. Perhaps a fundamentally new approach will render the problem obsolete someday. For now, and for at least the next few years, context management remains a practical concern. Antigravity chooses not to distract us with such details. There is no way to know where you sit in the context cycle, and therefore no way to take any corrective action even if you wanted to.
Opening a new chat compounds the confusion. The fresh session appears to inherit context from the previous one. Occasionally, continuity might be convenient, but more often than not, it is actively harmful. When I open a new chat, I want to do something entirely different. Perhaps an unrelated task. Perhaps a related task that demands a completely new set of instructions. The geniuses at Google apparently believe I am incapable of composing my own opening prompt, so they copy over the old material regardless. It is annoying. It is also deeply confusing. Does this mean the underlying context of the old chat has been carried forward as well, or is it merely a cosmetic echo? I checked the settings. I disabled what I believe controls the freshness of a new session. But how can I be certain when there is no indicator of context consumption, as previously discussed? I cannot. I simply have to trust that Google knows what it is doing, which brings us back to the central problem.
The inability to switch between different models on the fly, and between separate chats, barely warrants its own paragraph. It is a minor irritation stacked atop the larger structural failures, particularly given that multiple chats in the same window are already impossible.
I want to give Google some credit for trying. The Gemini 3.8 Flash model is genuinely excellent. It performs on par with rival offerings from competing firms while costing significantly less, making it a compelling choice for cost-conscious developers. Google has vast engineering talent. They understand artificial intelligence better than most organisations on the planet. They remain at the industry vanguard in numerous respects. So it is genuinely baffling that the IDE itself feels like an afterthought, a project receiving negligible investment or attention.
This would not be the first time Google neglected one of its own promising creations. There is an entire graveyard of ambitious projects that were launched with fanfare and quietly suffocated. Google Plus is the most obvious example, but the list stretches back years. I genuinely hope Antigravity IDE avoids that fate. I hope someone at Google reads this and takes action. After all, it is not 2023. It is 2026. Adding the missing features should not require a team of hundreds. It should not take more than a few well-crafted prompts typed into the very same chat window where they are already trying to do all their work at once.