Flaborea

earlier this week my team in google deepmind launched @antigravity! it's a new agentic code product, in IDE form, that's closely tuned to the behavioral modes of gemini 3 and makes the most of gemini's code capabilities. we use Antigravity heavily within the Antigravity team and across google deepmind. admittedly, i use AI code tools sparingly, since i learn more without heavy AI assistance, but i've personally found great uplift calling on the Antigravity agent for various research eng tasks, from bypassing the cruft in implementing RL infra or data pipelines, to rebasing a complicated experiment in the google monorepo after a coworker has refactored something big. normally these are high-stakes tasks with little room for error, and i'm only comfortable using the Antigravity agent for them if i *trust* the agent. as such, i specifically wanted to shout out our Artifacts feature: these are agent-generated markdown files where both the AI agent and the user can track the planning, progress, final results / findings, and more, across the agent's task. i grabbed the example pictures below from @kevinhou22's walkthrough of antigravity on youtube (linked in a follow-up tweet). besides this notion of user<->agent trust, i think these artifacts grant many more advantages w.r.t. UX and model capabilities. as models get stronger, we're increasingly using them for complex, long-horizon coding tasks, where it's a lot of scutwork to verify these tasks after an AI has (maybe incorrectly) done them, and where having an async agent set-and-forget these tasks doesn't feel like the most participatory or empowering mode of software development. on the other hand, within our Artifacts, i've personally found the implementation plan, for example, to be the perfect balance between having my own high-level supervision of the task vs. letting the agent do the work. the great thing is, *gemini is trained to adhere strongly* to the implementation plan and other planning/progress artifacts. if something is outlined correctly in a planning artifact, i know it will be done correctly, and if something is missing, i can easily catch it and course-correct the model before it executes. these features are one example of how the Antigravity team strives for forward-thinking design in agentic code development. this initial launch is only the beginning for us--the Antigravity team is fired up to keep shipping big improvements and creative new features for the product. in general i've found it incredible how a massive company like google / google deepmind has mobilized around the pace of frontier AI development, and internally there is a pervasive, ebullient impetus towards moving fast and shipping cool shit just like a startup. the future is bright! hope you enjoy Antigravity :)