Tech • IA • Crypto
Build agents with Gemini API (I/O Connect ‘26)
9.7/10TL;DR
Google DeepMind showcased rapid advances in Gemini models and new tools that simplify building AI agents and full-stack applications directly within AI Studio.
KEY POINTS
Rapid evolution of Gemini models
Google DeepMind has accelerated development from Gemini 1.5 through Gemini 2.5 to Gemini 3.5 Flash, now positioned as a preferred general-purpose model. Alongside proprietary systems, the company released open models like Gemma 4 and expanded into multimodal and media generation, including audio and text-to-speech systems. These models are increasingly used across industries, with adoption from firms such as GitHub, Figma, and Vercel.
Multimodal capabilities drive adoption
A key differentiator of Gemini is its ability to process and generate across formats, including text, images, audio, video, and code. Users can combine inputs such as URLs, speech, and system instructions, and receive outputs ranging from media files to executable function calls. This flexibility enables complex workflows and creative applications without switching tools.
AI Studio becomes full-stack development hub
AI Studio now supports building complete Node.js applications with automated server setup, integrated secret management, and access to the full npm ecosystem. Developers can prototype, test, and deploy applications without leaving the browser, reducing reliance on external infrastructure and simplifying development workflows.
Native Firebase and Workspace integrations
New integrations with Firebase provide built-in authentication, including Google sign-in, and automatic provisioning of Firestore databases. Connectivity with Google Workspace apps like Gmail and Docs enables developers to build tools that interact directly with personal or corporate data, eliminating complex configuration steps.
One-click mobile app generation
A major update allows developers to generate native Android apps using Kotlin and Jetpack Compose from natural language prompts. More than 1 million apps have reportedly been created within a month of launch. The system includes a browser-based emulator, zero setup, and one-click deployment to devices or app distribution channels.
Improved developer experience with AI assistance
Features such as focus annotation let users modify specific UI elements directly, while design previews provide multiple interface options during generation. Real-time prompt autocompletion helps overcome writer’s block, accelerating iteration and reducing friction in app design.
Enterprise-ready deployment options
Applications can be deployed directly to Google Cloud Run or hosted via Firebase without managing servers. These workflows support enterprise requirements while maintaining simplicity for smaller teams and individual developers.
Introduction of the Interactions API
The newly general-available Interactions API provides a unified interface for both models and agents. It supports multimodal inputs, tool usage, and a new step-based interaction model that replaces traditional chat turn structures, making it easier to manage reasoning, tool calls, and outputs.
Shift toward agentic workflows
The API reflects a broader move toward goal-driven AI agents capable of executing multi-step tasks. Features include asynchronous execution, server-side context management, and support for long-running processes, enabling more complex automation beyond simple prompts.
Managed agents and remote environments
Developers can run agents in isolated cloud sandboxes with persistent environments, allowing tasks like installing dependencies or analyzing data without local setup. These environments support integration with repositories, storage systems, and private data sources.
Tool integration and security enhancements
The platform allows combining first-party tools like Google Search with custom functions. A proxy system ensures API credentials remain secure by injecting them at runtime, preventing direct exposure within agent environments.
Simplified agent creation via files
Agents can now be defined using markdown instructions and lightweight scripts instead of complex code loops. Capabilities are modular, with “skills” enabling tasks such as text-to-speech, research, or media generation, significantly lowering the barrier to entry.
CONCLUSION
Google DeepMind is positioning Gemini and AI Studio as an integrated ecosystem for building, deploying, and scaling AI agents, aiming to reduce technical overhead while expanding the scope of what developers can automate.