Each development in mannequin intelligence for coding has inspired us to rethink how growth must be achieved. The Built-in Growth Setting (IDE) of at present is a far cry from the IDE of only a few years in the past. Gemini 3, our most clever mannequin, represents a step-change for agentic coding, and requires us to consider what the subsequent step-change of an IDE must be.
Immediately, we’re introducing Google Antigravity, our new agentic growth platform. Whereas the core is a well-recognized AI-powered IDE expertise with the perfect of Google’s fashions, Antigravity is evolving the IDE in direction of an agent-first future with browser management capabilities, asynchronous interplay patterns, and an agent-first product kind issue that collectively, allow brokers to autonomously plan and execute complicated, end-to-end software program duties.
Why We Constructed Antigravity
We would like Antigravity to be the house base for software program growth within the period of brokers. Our imaginative and prescient is to in the end allow anybody with an concept to expertise liftoff and construct that concept into actuality. From at present, Google Antigravity is accessible in public preview at no cost, with beneficiant price limits on Gemini 3 Professional utilization.
With fashions like Gemini 3, we now have began hitting the purpose in agentic intelligence the place fashions are able to operating for longer intervals of time with out intervention throughout a number of surfaces. Not but for days at a time with out intervention, however we’re getting nearer to a world the place we interface with brokers at increased abstractions over particular person prompts and gear calls. On this world, the product floor that allows communication between the agent and person ought to feel and appear completely different – and Antigravity is our reply to this.
Core Tenets
Antigravity is our first product that brings 4 key tenets of collaborative growth collectively: belief, autonomy, suggestions, and self-improvement.
Belief
Most merchandise at present stay in considered one of two extremes: both they present the person each single motion and gear name the agent has made or they solely present the ultimate code change with no context on how the agent received there, and with no simple solution to confirm the work. Neither engenders person belief within the work that the agent undertook. Antigravity gives context on agentic work at a extra pure task-level abstraction, with the required and enough set of artifacts and verification outcomes, for the person to realize that belief. There’s a concerted emphasis for the agent to completely assume via verification of its work, not simply the work itself.
In a dialog with an agent in Antigravity, the person sees device calls grouped inside duties, monitoring excessive stage summaries and progress alongside that activity. Because the agent works, it produces Artifacts, tangible deliverables in codecs which might be simpler for customers to validate than uncooked device calls, reminiscent of activity lists, implementation plans, walkthroughs, screenshots, and browser recordings. Brokers in Antigravity use Artifacts to speak to the person that it understands what it’s doing and that it’s completely verifying its work.
play_arrow
View the Agent’s activity listing,evaluate the implementation plan post-research and pre-implementation,or scan the walkthrough at completion.
Autonomy
Immediately, essentially the most intuitive product kind issue is working synchronously with an agent embedded inside a floor (an editor, browser, terminal, and many others). That’s the reason Antigravity’s main “Editor view” is a state-of-the-art AI-powered IDE expertise, with Tab completions, in-line Instructions, and a completely functioning agent within the aspect panel.
That being stated, we’re transitioning to an period, with fashions like Gemini 3, when brokers can function throughout all of those surfaces concurrently and autonomously:
play_arrow
Autonomously,an Antigravity Agent writes code for a brand new frontend function,makes use of the terminal to launch localhost,and actuates the browser to check that the brand new function works.
We additionally imagine brokers deserve a kind issue that exposes this autonomy optimally and permits customers to work together with them extra asynchronously. So, along with the IDE-like Editor floor, we’re introducing an agent-first Supervisor floor, which flips the paradigm of brokers being embedded inside surfaces to at least one the place the surfaces are embedded into the agent. You may consider it like a mission management for spawning, orchestrating, and observing a number of brokers throughout a number of workspaces in parallel.
play_arrow
Consumer spawned an Agent to do background analysis in a special workspace whereas specializing in a extra concerned activity within the foreground,utilizing the Inbox and aspect panel within the Agent Supervisor to be notified of progress.
We determined to not attempt to squeeze each the asynchronous Supervisor expertise and the synchronous Editor expertise right into a single window, quite optimizing for instantaneous handoffs between the Supervisor and Editor. Antigravity is designed to be forward-looking, intuitively bringing growth to the asynchronous period as fashions like Gemini proceed to quickly get smarter.
Suggestions
A core failing of a remote-only kind issue is the shortcoming to iterate simply with the agent. Agentic intelligence has certainly improved considerably, however it’s nonetheless not excellent. An agent having the ability to full 80% of the work must be helpful, but when there is no such thing as a simple manner to offer suggestions, then it turns into extra work than profit to resolve the remaining 20%. Consumer suggestions permits us to not deal with an agent as a black-or-white system that both is ideal or unhelpful. Antigravity is beginning with native operation, permitting intuitive async person suggestions throughout each floor and Artifact, whether or not it’s Google-doc-style feedback on textual content Artifacts or select-and-comment suggestions on screenshots. This suggestions can be robotically integrated into the agent’s execution with out requiring you to cease the agent’s course of.
play_arrow
An instance of suggestions on textual Artifacts reminiscent of an implementation plan and an instance of suggestions on a visible Artifact reminiscent of a screenshot taken by the Agent.
Self-improvement
Antigravity treats studying as a core primitive, with agent actions each retrieving from and contributing to a data base. This data administration permits the agent to be taught from previous work. This might be necessary express data reminiscent of helpful snippets of code or derived structure, however is also extra summary, such because the sequence of steps taken to efficiently full a selected subtask.
play_arrow
Agent studying from work and suggestions to generate and leverage data gadgets,viewable from the Agent Supervisor.
Strive Antigravity Immediately
We imagine Antigravity’s product kind issue represents the subsequent elementary step operate in agent-assisted growth. Thus, our aim is to channel it into the perfect product providing potential for finish customers. In at present’s public preview:
- Google Antigravity for people at no cost
- Compatibility with MacOS, Linux, and Home windows
- Entry to Google’s Gemini 3, Anthropic’s Claude Sonnet 4.5 fashions, and OpenAI’s GPT-OSS throughout the agent, providing builders mannequin optionality 1
Study extra concerning the options of Antigravity in our docs and skim extra about how Antigravity can help you by shopping numerous use circumstances. We may have new options dropping often, so hold up-to-date with this weblog and socials at X, LinkedIn, and YouTube.
Expertise liftoff in 3… 2… 1…
