Google AI Updates: May 21, 2026

1. Antigravity 2.0 ships as a standalone agent desktop with subagents, hooks, and async task primitives

Google. Google launched Antigravity 2.0 on day two of I/O as a standalone desktop application positioned as a “central home for agent interaction,” with new core primitives for subagents, hooks, and asynchronous task management so a developer can run one agent coding while another generates brand assets in parallel. The release reframes Antigravity from a Gemini-CLI replacement into an agent-first development surface that compresses what Google calls “multi-day engineering efforts” into hours or minutes by orchestrating fleets of cooperating subagents. Antigravity 2.0 ships alongside an Antigravity CLI for headless terminal-first use and an Antigravity SDK that exposes the same agent harness used inside Google’s own products and is co-optimized for Gemini, so teams can host customized agents on their own infrastructure. Source

2. Managed Agents land in the Gemini API as one-call remote Linux sandboxes

Google. Google added Managed Agents to the Gemini API, where a single call provisions a remote Linux environment in which an agent reasons, plans, calls tools, manages files, and browses live web data through the Antigravity harness. The default agent is built on Gemini 3.5 Flash and exposed through both the Interactions API and Google AI Studio, with developers able to extend agents through custom instructions and AGENTS.md/SKILL.md markdown files and then register them as named, reusable agents. The product slots directly against Claude Managed Agents and AWS Bedrock AgentCore, with the differentiation being that the entire harness and sandbox lifecycle are owned by Google rather than the customer. Source

3. Gemini for Science launches three tools spanning hypothesis generation, parallel code search, and literature insights

Google. Google unveiled Gemini for Science, a suite of three experimental tools on Google Labs aimed at research workflows: Hypothesis Generation, built on Co-Scientist, runs a multi-agent “idea tournament” that generates, debates, and evaluates hypotheses with cited claims; Computational Discovery, built on AlphaEvolve, generates and scores thousands of code variations in parallel so scientists can test approaches in fields like solar forecasting in minutes rather than months; and Literature Insights, built on NotebookLM, searches scientific literature and structures results into side-by-side comparison tables. Google also previewed Science Skills, a bundle that integrates 30+ life-science databases so structural bioinformatics and genomic analyses can be run “in minutes” from a single agent prompt. The release is the most direct Google bet so far that LLM agents become a primary interface for empirical research, not just literature search. Source

4. AI Studio adds native Android app building with one-click Play Internal Test Track publishing

Google. Google AI Studio added native Android app building, letting developers select “Build an Android app” and generate native Kotlin/Jetpack Compose applications directly from prompts. The flow integrates with Play Console: builds can be previewed in an in-browser Android Emulator or installed on devices via ADB, then published to the Google Play Internal Test Track with a single click, with the first two apps getting free Google Cloud quota and no credit-card requirement. The feature is the most prompt-to-store loop Google has shipped, and it lands the same week that several analyses framed AI Studio’s Android pipeline as a potential disintermediation of the conventional App Store/Play Store middle layer for utility-class apps. Source

5. Build with Gemini XPRIZE opens with a $2M hackathon prize pool

Google. Google announced Build with Gemini XPRIZE, framed as the largest hackathon prize pool to date at $2 million, inviting developers to build applications addressing urgent global challenges on top of Gemini and the new agent stack. The contest channels traffic toward the new Antigravity 2.0 and Gemini for Science surfaces, with submission guidance pointing developers at the Managed Agents API, Antigravity SDK, and the new Skills format. It is the highest-dollar XPRIZE-branded developer competition Google has run and signals that the company expects the next wave of Gemini adoption to be developer-led rather than consumer-led. Source