AI
Managed Agents Ship with a Sandbox
Logan Kilpatrick announced managed agents in the Gemini API. Developers add custom instructions, tools, and data through a single API call. Every managed agent comes paired with a remote Linux environment hosted by Google. The team at Stitch used managed agents to let users import a design system from a GitHub repo. The agent analyzes the codebase and generates a design.md file. Early access customers include Ramp, Resemble AI, and Clippy.
AI Studio Generates Android Apps
Paige demonstrated building an Android app directly in AI Studio. No SDKs or local environment are needed. The app is built entirely in Kotlin. Users can preview the app with an Android emulator inside AI Studio. Publishing to the Google Play Store test track is a few clicks from the same interface. A new mobile app for AI Studio rolls out in a few weeks.
Anti-Gravity 2.0 Adds Sub-Agents and Cron
Anshul Ramachandran introduced anti-gravity 2.0, a desktop application for orchestrating multiple agents. Dynamic sub-agents spin up specialized helpers like a QA or data science agent that work in parallel. Scheduled tasks use standard cron expressions to put agents on autopilot. Kevin Hou then used the anti-gravity CLI to fine-tune Gemma 4 on stage. He fed a stack trace to the agent, which wrote the LoRA training code and deployed the custom model to his laptop.
Android CLI and Knowledge Base Cut Token Use
Florina Muntenescu and Adarsh Fernando showed the Android CLI built into anti-gravity. It includes an Android knowledge base and open-sourced Android skills. Internal testing showed agents using these tools use 70% fewer tokens and complete tasks up to three times faster. The agent built a display glasses UI using Jetpack Compose Glimmer from the Android XR SDK. It also used Android Studio to look up dependency versions and analyze files for issues.
Modern Web Guidance and Chrome DevTools for Agents
Yuna announced modern web guidance, a set of skills for AI agents that covers over 100 use cases for the latest web platform features. For web development tasks, guided coding improves pass rates by an average of 37 percentage points. Matias then demonstrated Chrome DevTools for agents. The agent runs a Lighthouse audit with a new Agentic browsing category that validates WebMCP tool registrations and LLM.txt files. The agent reads the report, attempts a fix, and reruns the audit.
HTML Elements Inside Canvas
Matias and Yuna revealed the new HTML and Canvas API. It renders real DOM elements directly into a canvas environment. Those elements are searchable, accessible, selectable, translatable, and interact with browser features like autofill. The API is available as an origin trial now.
Notable Quotes
I didn’t write any orchestration logic. I simply define the skills and tools in Markdown files and the agent does the rest. Logan Kilpatrick · ▶ Watch (6:48)
In minutes, we used anti-gravity to go from no code to fine-tuning Gemma 4 on our own custom use case. And if all that seems a little bit too simple, that is the point. Kevin Hou · ▶ Watch (20:40)
We saw in our internal testing, agents leveraging Android CLI, skills, and knowledge base use about 70% fewer tokens and complete tasks up to three times faster. Florina Muntenescu · ▶ Watch (29:17)
The new HTML and canvas API now does the impossible. With this, you can now integrate real DOM elements directly into your canvas environment. Matias · ▶ Watch (54:33)
Key Takeaways
- Managed agents in the Gemini API include a remote Linux sandbox per agent.
- AI Studio generates Kotlin Android apps and deploys to Play Store.
- Anti-gravity 2.0 supports dynamic sub-agents and cron-based scheduling.
- Modern web guidance improves agent coding pass rates by 37 percentage points.
- Chrome DevTools for agents closes the feedback loop with automated Lighthouse audits.