FR version is available. Content is displayed in original English for accuracy.
Advertisement
Advertisement
⚡ Community Insights
Discussion Sentiment
73% Positive
Analyzed from 1201 words in the discussion.
Trending Topics
#antigravity#google#security#tui#agent#code#agents#gemini#https#cli
Discussion (32 Comments)Read Original on HackerNews
https://antigravity.google/blog/introducing-google-antigravi...
Discussion: https://news.ycombinator.com/item?id=48196838
Reasonix is not great on all fronts, but it shows how much money you've spent at all times, it's fast, it's not a direct copy of Claude Code, and it is designed specifically to maximize DeepSeek caching, so it's free real estate...like a buck a day to use DeepSeek models hard.
MiMo Code is just cute as hell. It's got (mostly tasteful) animations and emojis and such, it's fun to use. It also is not a direct copy of Claude Code. It almost feels like a GUI app. MiMo Pro is an underappreciated model, too, at least for some classes of problem. It's very good at security vulnerability research, and very cheap. It's slower than DeepSeek and more easily confused by tools, but it's uncannily good at reading code and identifying security bugs (rough parity with Opus 4.8/GPT 5.5 at more than an order of magnitude lower price); and with fewer false positives than most models, including the frontiers. And, to keep it on topic, Antigravity refuses security work of any sort, so it's unfit for purpose. Gemini 3.6 Flash is pretty good at security work via the API, so it's not the model rejecting, it's the agent, which is extra annoying.
I also don't really get why TUI's have become such a big deal, other than maybe nerd bait. The only time I want a TUI is when I'm SSH'd into a machine, which isn't very often? TUI's have so many downsides. Selecting text is inconsistent, copy paste is inconsistent, mouse behaviors.. it's all kind of jagged.
First up - I completely agree that parts of the terminal experience are decidedly lackluster, you can make them "fine" but the general experience doesn't align with consistent OS principles users are familiar with, and I agree that good mouse support, expected copy and paste, selection behaviors, adding files/attachments to prompts, etc... all could be better (and actually are mostly better on decent modern terminals, as an aside)
But... And this is the big but... LLMs are basically text IO systems. And terminals are designed from the ground up to allow text IO to control the entire system, including the TUI (ex - multiplayer agent tmux is a fun demo).
It's a genuinely powerful pairing, and basically ALL of the non-TUI systems basically end up just embedding a terminal [cough* xterm *cough] and becoming a non-standard TUI themselves, anyways.
So TUIs have become a big deal because it's the obvious and easy way to give an LLM "computer use" skills, without spending a fortune on tokens doing graphical processing for GUIs, which are geared towards humans and not agents, and are generally less powerful anyways.
And not sandboxing your agents seems insane.
Well, that's your need and it is fine. For me, I want TUIs all the time for everything. They blend really nice with... well, all the other TUI tools I already use all the time.
The current CLI version 1.1.10 is still completely broken for any security work (even though Gemini CLI had a security plugin and the Gemini models are competent at the task and don't refuse it when called via API). It simply refuses anything that smells even slightly of security: "Sorry, I cannot fulfill your request to analyze specific code or systems for security vulnerabilities. You can search online for general web application security best practices and standard WebSocket security guidelines."
Classic Google. Kill a pretty good product, replace it with a defective one. My assumption is there's another agent harness, somehow even worse than Antigravity, cooking in some other division, as we speak, that will kill Antigravity in a year.
Antigravity is a whole different product, the only way I use it currently is to abstract the OAuth mechanism so we can leverage Google subscriptions within our own agents.
Security audits are the one thing everyone needs to be doing with AI, now, even if you're not doing any other work with AI. If you hate LLMs you still need to defend against them. And, the US providers seem to all be racing to make that capability inaccessible to anyone who isn't at a Fortune 500 and able to use their "cyber" versions.
Because it's valuable. I'd be surprised if the big providers didn't go even further with this idea. Need your agent to write some C++? Gotta upgrade to the "system developer" package (a $100/mo add-on to your subscription). SQL? You'll need the DBA package. Working with hardware/firmware (or anything that might be construed as reverse engineering or hardware hacking)? You'll need a verified computer engineering account for that.
Coordination overhead and duplicated work are usually where these systems start to fall apart for me.
https://github.com/kunchenguid/firstmate
Not affiliated, just started using a few weeks back and am pleasantly surprised how well it works even with rather dumb agents.
Either way good for Google but I do not expect this to gain a lot of traction, it has been almost months now since I've heard someone talk positively about the Google models and other AI offerings.
It's too little, too late.