How We Used Gemini to Build Google I/O 2026
When Google told Gemini CLI developers they had until June 18 to switch to a new command-line interface called Antigravity, it was the most concrete signal in months about where the company is actually putting its developer relations weight.
The forced migration announcement — buried in a May 19 blog post alongside pricing changes and benchmark updates — was the part of Google's I/O 2026 production story that went almost entirely uncovered. Developers who built workflows around Gemini CLI are now being asked to learn a new CLI with a different mental model, different commands, and a different product identity. The transition is not optional. For a developer audience, this is not a footnote — it is the story.
Google published the rest of the I/O production pipeline on June 1: a production stack connecting AI Studio, Gemini Canvas, Antigravity, Coral NPU, and Lyria 3 Pro, all orchestrated to build the apps, music, and visual identity for the company's flagship event. The post confirmed what Google had connected and was running at scale. That is a different kind of claim than a model release. And it is one that only Google is in a position to make about itself.
The self-attestation problem is the spine of this story. When a company uses AI to produce the showcase for its own AI announcements, it eliminates any independent vantage point from which an outside observer can verify what the technology actually contributed. Google is not unique in this bind — the entire AI industry exists inside a credibility problem, where the organizations best positioned to prove AI's value are the ones with the most incentive to oversell it. But the June 1 post made that problem visible in a specific and uncomfortable way.
The orchestration stack Google described is real, and it represents a genuine architectural commitment. The company built a custom tool inside AI Studio to test Nano Banana — Google's image generation model, later rebranded as Gemini 3 Pro Image — frame outputs at scale, ensuring pixel-perfect consistency before generating sequences. That is the kind of infrastructure investment that signals long-term priority, not a one-off experiment. Antigravity, the agent-first development platform that launched at I/O with a desktop app, CLI, SDK, and enterprise support through Google Cloud, was used to build the Antigravity Coffee Co. app distributed to attendees. Coral NPU hardware handled real-time inference for the jellyfish-tracking music generation that powered the pre-show. Gemini 3.5 Flash became the default model powering the agent harness, and Managed Agents launched in the Gemini API, letting developers spin up agents that reason, use tools, and execute code in isolated Linux environments with a single API call.
Independent testing of Antigravity's multi-agent architecture paints a more complicated picture than Google's presentation suggests. In hands-on comparisons with Cursor and Claude Code on real Python projects, Antigravity consistently outperformed on complex autonomous tasks — the kind where a tool plans, executes, and self-corrects across multiple steps. It underperformed on speed and debugging depth, where Claude Code's reasoning quality and Cursor's raw velocity held advantages. The multi-agent approach matters: Antigravity can spawn dynamic subagents that work in parallel, run tests automatically, and iterate without human intervention. That is genuinely different from what Cursor or Claude Code offer. But it is not a clean win.
Former Google senior staff engineer Rui Diao, writing in his newsletter The Signal, called the Antigravity versus AI Studio overlap a political problem between two product teams and noted that Google now has what may be the most confusing AI brand portfolio in the industry — Imagen, Veo, Flow, AI Studio, Antigravity, Gemini, Nano Banana, Gemma. The I/O blog post did nothing to reduce that confusion. It described a production process, not a product.
That is both the limitation and the opportunity of the post. Google showed its work in enough detail to be informative about its engineering priorities, but the frame was fixed: this is what we built, and it worked. There is no outside perspective on whether the AI actually saved time, whether the production quality exceeded what a conventional pipeline would have produced, or whether any of the claimed efficiencies are replicable by teams outside Google with comparable resources. The post's own conclusion — that when AI is used well, you stop noticing it — was offered without anyone being in a position to offer a contrary view.
The credibility problem here is not unique to Google. When Anthropic publishes benchmarks, the world can test them. When OpenAI releases a model, developers build on it and report back. When Google publishes a blog post describing how Gemini built Google I/O, there is no equivalent feedback mechanism. The company that made the product is the company that produced the evidence. The verification loop closes before it opens.
Builders who want to understand what Google's AI stack actually offers for production workflows have the company's own documentation, third-party hands-on testing from outlets like Plain English and Lushbinary, and the competitive comparison data against Cursor and Claude Code. The Antigravity platform is real and the multi-agent architecture is a genuine architectural bet. The forced migration by June 18 is real and has immediate consequences for every developer who built on Gemini CLI. The self-attestation problem is also real, and it means that every efficiency claim Google makes about its own internal use of these tools should be treated as a direction signal, not a verified result.
The most honest thing to say about Google's I/O production post is that it describes a real product, built by real engineers, solving real production problems — and that nobody outside Google is positioned to confirm whether the version of that story in the blog post is the complete version. That is not a novel problem in technology journalism. It is just unusually visible in this case, because the company used the product to build the announcement of the product, and then wrote the announcement itself.
What builders should take from the post: the orchestration stack is the story, not any individual model. AI Studio, Antigravity, and Gemini 3.5 Flash together represent Google's current answer to the question of how production AI workflows actually get built. The June 18 migration deadline is the concrete near-term consequence. Whether that answer is better than what Cursor, Claude Code, or the next generation of developer tools will produce depends on problems that have not been solved yet — and claims that have not been independently verified.