February 2026 Summaries

The Agent Development Kit (ADK) is an open-source framework designed to enhance AI agents by enabling them to interact with various systems, facilitating a transition from mere conversation to actionable tasks. ADK supports developers in creating, testing, and deploying complex workflows with integrations that allow agents to manage code repositories, trigger workflows, query databases, and more. The framework has recently expanded through partnerships with major developer platforms, offering a broad array of third-party tools and integrations. These integrations cover areas such as code development, project management, databases, memory, observability, external connectors, AI models, payments, speech, and email, providing a comprehensive suite for AI agents to operate effectively in real-world scenarios. With a focus on a unified developer experience, ADK simplifies the integration of these tools through a plugin architecture, allowing developers to enhance their agents' capabilities with minimal code modifications.

Google has announced new updates to its Google AI Edge Gallery app, now available on both Android and iOS platforms, enhancing on-device AI interactions through efficient function-calling. This development showcases the FunctionGemma model, which efficiently translates natural language into function calls on mobile devices, enabling offline capabilities without the need for server interaction. Key features include Mobile Actions and Tiny Garden, two demos that highlight how natural language commands can be seamlessly executed on-device, such as creating calendar entries or interacting with a virtual garden, all with a compact 270M parameter model. The app also includes benchmarking tools to allow users to evaluate the performance of these models on their own devices, maintaining high-speed operation even on mobile hardware. By expanding to iOS, Google ensures that its AI-powered tools offer privacy and reliability across different mobile ecosystems, allowing developers to explore and customize AI interactions for personal use cases.

Gemini's new Canvas feature in the Gemini web app enables the creation of immersive, interactive 3D experiences in extended reality (XR), making it accessible to developers and users without extensive skill requirements. The Android XR team demonstrated this capability by developing a biology simulation that visualizes blood cells at a sub-cellular scale using WebGL, Three.js, and WebXR APIs. This innovation allows for rapid prototyping and exploration of interactive 3D concepts, facilitating the transformation of creative ideas into functional experiences with speed and ease. Users can begin creating their own XR content using a Samsung Galaxy XR headset and Gemini 3 Pro, utilizing pre-built XR Blocks Gems or developing their own with advanced perception capabilities. The process is streamlined through Gemini’s tools, enabling users to build and iterate on their ideas, resulting in shareable web links of their XR creations. Gemini encourages experimentation with starter prompts and advanced challenges, such as making interactive environments or creating lifelike sculptures, promoting creativity and collaboration within the XR community.

Google I/O is set to return on May 19–20 with an online event showcasing the company's latest AI advancements and updates across various products, including Gemini, Android, Chrome, and Cloud. Participants can expect keynote speeches from Google leaders, insightful panel discussions, and product demonstrations highlighting technological innovations. The event will also feature Dialogues sessions where prominent thinkers and leaders will discuss the impact of AI on the future. Interested individuals can register on io.google to receive updates and tune in for the livestream beginning at 10 am PT on May 19.

Conductor, an extension for the Gemini CLI, enhances AI-assisted engineering by integrating context-driven development into the terminal through version-controlled markdown files. Its new Automated Review feature adds a "verify" step to the development lifecycle, generating reports on code quality and compliance to predefined guidelines. This feature includes code review, plan compliance checks, guideline enforcement, test-suite validation, and basic security reviews, focusing on identifying issues such as race conditions, logic errors, and security vulnerabilities. The findings are categorized by severity, providing developers with actionable intelligence to address issues, ensuring a supervised AI development process. This approach allows developers to maintain high-level architectural oversight while the AI handles labor, contributing to a safe, predictable, and sound development environment.

Gemini CLI has introduced a new feature to enhance the configuration of its extensions, aiming to streamline and secure the setup process for users. This update involves a structured approach to extension settings, allowing authors to specify required information such as API keys or project identifiers, which are then prompted during installation. This automated setup ensures sensitive data is securely stored in the system keychain and not in plain text, improving both ease of use and security. The Gemini CLI also offers a command-line tool, 'gemini extensions config,' for managing and updating settings post-installation, with support for global and project-specific configurations. A practical example is the AlloyDB extension, which necessitates several specific inputs to function, now simplified through guided setup. The update aims to provide a first-class user experience by eliminating manual configuration errors and enhancing visibility into current settings, exemplified by the ability to list and verify active configurations easily. This feature is available from version 0.28.0 onward, encouraging developers to adopt these practices for better extension management.

In September 2025, the Data Commons Model Context Protocol (MCP) server was introduced to standardize AI agents' interaction with Data Commons data, initially requiring local Python environments. This posed challenges in high-security environments and scalability, particularly for developers wishing to publish query agents. To address these issues, a hosted MCP service on Google Cloud Platform was launched, eliminating the need for local setups and allowing users to connect seamlessly to datacommons.org. The service enables analysts and developers to use natural language queries to draw insights from trusted data sources, facilitating the creation of AI agents tailored to specific needs. Users of the Gemini CLI extension benefit from automatic updates, while others need a free Data Commons API key to connect. The hosted service is specific to datacommons.org, and custom instances require a separate server setup.

The announcement of the public preview of the Developer Knowledge API and the Model Context Protocol (MCP) server marks a significant advancement in providing AI-powered developer tools with access to accurate and current Google developer documentation. The Developer Knowledge API aims to serve as the programmatic source of truth, offering comprehensive coverage and timely updates of documentation from sources like firebase.google.com and developer.android.com, re-indexing within 24 hours of updates to ensure freshness. The accompanying MCP server, an open standard, allows AI assistants to access this documentation safely and efficiently, enabling features such as implementation guidance and troubleshooting. This preview encourages developers to integrate these tools into their workflows, with plans for future enhancements to include structured content and broader documentation coverage.

FunctionGemma is a small language model designed for efficient API call translation on edge devices, and its fine-tuning process can be enhanced by using Google Tunix on TPUs, as explored in this tutorial. Tunix, part of the JAX AI Stack, facilitates post-training techniques such as supervised fine-tuning and model distillation across various large language models, including Gemma and LLama. By employing the LoRA method for supervised fine-tuning, the tutorial demonstrates how to set up the FunctionGemma model on free-tier Colab TPU v5e-1, showcasing Tunix's ability to achieve high TPU utilization and significantly improve model accuracy with minimal training overhead. The process involves downloading model weights and datasets via Hugging Face, using JAX for parallelism, and creating a custom dataset class to feed training data into Tunix. The training process is completed using the PeftTrainer, highlighting Tunix's potential to drive qualitative improvements and making it an invaluable tool for developers refining LLMs for specific applications.