Video: "I Tried Google's New Antigravity 2.0 (Full Build)" by Julian Goldie on YouTube.
What Antigravity 2.0 actually is
Google Antigravity 2.0 is the second major release of Google's autonomous coding environment, updated significantly at Google I/O 2026. Unlike a code assistant you interact with in a chat window, Antigravity operates through sub-agents — one plans the architecture, one writes the code, one runs tests, one handles error diagnosis. These run in parallel rather than in sequence, which is the main reason it's faster than tools that handle each step separately.
It integrates directly with Google's AI stack: Gemini handles the reasoning, Vertex AI provides the infrastructure, and the output connects naturally into Google Cloud and Firebase for deployment. A basic tier is available free through Google Workspace, though the more capable agentic features — particularly the multi-agent coordination and longer autonomous runs — sit behind paid tiers. Worth noting: it's still early, and the feature set shifted between the I/O announcement and the version Julian tested.
The full build: what Julian actually asked it to do
Julian started with a plain English brief: build a functional web application for a specific use case, starting from scratch, with no codebase provided. Antigravity 2.0 ran through the build in several phases. First, it produced a spec — file structure, technology choices, core components — before writing any code. That planning phase is one of the more useful differences from simpler AI coding tools, which often jump straight to writing and create structural problems you spend hours unpicking later.
With the spec confirmed, the build agents wrote the main application files while the testing agent started defining test cases. The two ran simultaneously, which compressed the overall time. When tests failed — and some did — the error diagnosis agent identified the issue and proposed a fix, which the build agent then applied. The loop ran several iterations before producing something that passed the test suite.
Where it was genuinely impressive
The planning phase was better than expected. Most AI coding tools treat architecture as an afterthought; Antigravity 2.0 laid out a coherent project structure before touching the code, which meant the eventual output was organised rather than a jumble of files that work but are difficult to maintain.
Error recovery also held up better than previous AI coding tools. The diagnosis was usually correct on the first try, and the fix didn't introduce new errors at the rate you'd expect from models that are essentially guessing at the problem. In practice, this cut the number of manual debugging sessions needed significantly. For straightforward applications — business tools, simple data dashboards, internal automation — it genuinely did most of the build autonomously.
Where it still needs a human
Design is the most obvious gap. Antigravity 2.0 produces functional output, not polished output. The UI it generates works but looks generic — standard component libraries, default spacing, nothing that reflects a specific brand or customer expectation. If the end product needs to look good, a designer still needs to review and rework the front end.
Business logic is the other limitation. Antigravity doesn't know your specific requirements, your existing data structure, or your users' actual behaviour. It builds what you describe in the brief, which means the quality of the brief determines the quality of the output. Vague briefs produce vague applications. And for complex integrations — connecting to existing APIs, matching specific authentication flows, handling edge cases in your particular business logic — the autonomous build still hits dead ends that need a developer's attention.
Where this connects to NordSys
Tools like Antigravity 2.0 are genuinely changing how custom software gets built — faster first drafts, less manual scaffolding, more time spent on the parts that actually require human judgement. We use AI coding tools including Claude Code, Codex and Antigravity to build applications and automations for UK businesses, handling everything from brief to deployment. If you have a specific build in mind, we can tell you honestly how much AI can do on its own and where it still needs experienced eyes.
See our Programming service →