May 2026 Summaries

The introduction of the Google Pay & Wallet Developer MCP server aims to enhance the integration process of Google Pay & Wallet APIs for developers by leveraging AI within their IDEs. By addressing the lack of real-time context and updates on Google Pay & Wallet accounts, the Model Context Protocol (MCP) provides tools for AI development assistants to facilitate smoother, error-free integration directly from the development environment. Key features include instant access to documentation, real-time account and integration details, on-the-fly validation of Wallet pass JWTs or JSON definitions, performance monitoring, and integration management. This setup reduces the need for context switching and enhances development efficiency by offering AI-generated code and answers based on the latest documentation and account data. The MCP server aims to reduce friction in the setup process, accelerate development, and improve accuracy, with future plans to expand its tools and capabilities based on user feedback within the developer community.

Large Language Models (LLMs) can benefit from structured reasoning training, as demonstrated in the Google Tunix Hack: Train a model to show its work hackathon on Kaggle. The event challenged developers to enhance non-reasoning models into general reasoning ones using limited computational resources, resulting in over 11,000 participants and 300 high-quality submissions. The winning techniques included innovative combinations of supervised learning, preference optimization, and reinforcement learning, such as the G-RaR (Rubric-Based Reinforcement Learning) which trains models to produce structured reasoning by using a rubric-based reward system. Other notable approaches included Pinocchio-1B and IDEA-E, which focused on structured reasoning through stages of fine-tuning and reinforcement learning, employing various reward systems to enhance logical deduction and prevent premature guessing. These efforts showcased that structured reasoning can significantly improve LLMs' performance across various domains, including medical, legal, and robotics, even with limited computational resources, democratizing the ability to train such models using publicly available recipes and tools like Tunix and Kaggle TPUs.

Google Pay is advancing into the era of agentic commerce with several new features and tools aimed at enhancing the developer platform and optimizing the checkout experience. These updates include compatibility with the Universal Commerce Protocol (UCP) using existing infrastructure, the introduction of the Google Pay & Wallet Developer MCP server to streamline AI integration, and dynamic callbacks for Android to facilitate a 1-click Express checkout. Enhancements also extend to social apps, offering seamless payments across various platforms. New tools provide transparency in transaction routing, including cardFundingSource signals and eftpos routing in Australia, while cross-device authentication aims to reduce friction during desktop checkouts by leveraging mobile-first authentication flows. These innovations are designed to support businesses in managing processing costs more effectively and ensuring the continuity of recurring transactions with lifecycle notifications for payment tokens.

Express checkout with Google Pay for Android native apps allows developers to utilize users' stored credentials from Google Wallet to simplify the checkout process. The implementation introduces dynamic callbacks such as onPaymentDataChanged and onPaymentAuthorized, enabling developers to update shipping options, taxes, and total prices dynamically within the Google Pay sheet without closing it. This feature is available from play-services-wallet:20.0.0 onwards and facilitates an "Express Checkout" experience by moving the Google Pay button to Product Detail or Cart pages, thereby improving conversion rates. Developers can handle transaction authorization and retries directly within the Google Pay interface, enhancing the checkout funnel by providing real-time updates on shipping and tax information based on user interactions. Dynamic callbacks align the Android Google Pay developer platform with its web capabilities, offering a streamlined, accurate, and efficient payment experience.

Google has announced the expansion of its Gemini AI capabilities to Google Home, transforming it into a full-stack AI offering that empowers service providers and hardware manufacturers to create monetizable, proactive home services. This development leverages Google Home APIs, enabling devices to deliver enhanced features such as camera intelligence for contextual notifications, the "Ask Home" function for complex inquiries, and "Home Brief" for comprehensive daily activity summaries. Service providers, including AT&T, can integrate these features into their services, enhancing security and user experience. Google is also introducing the Gemini built-in Program to facilitate the development of AI-native hardware, offering validated reference designs to expedite the creation of smart cameras and speakers. This initiative marks a significant expansion of the Google Home ecosystem, providing developers access to a comprehensive stack for building intelligent home solutions.

Agent Development Kit (ADK) has introduced version 0.1.0 for Kotlin and a specialized library for Android, expanding its open-source framework for developing AI agents. ADK for Kotlin facilitates the creation and execution of agentic workflows in backend projects, while ADK for Android allows AI agents to operate directly on devices using local large language models (LLMs), enhancing both privacy and efficiency. The framework is designed to manage complex task orchestration, context handling, and error management between cloud and edge environments, providing flexibility in choosing between on-device and cloud models. Key features include hybrid orchestration, on-device sequential agents, and local retrieval capabilities, exemplified by an in-app trip assistant that seamlessly integrates cloud and on-device functionalities. The ADK for Kotlin allows developers to equip agents with specific tools, demonstrated through the creation of agents inspired by "The Hitchhiker's Guide to the Galaxy," highlighting the ease of integrating advanced AI functionalities into applications. The release marks the beginning of a broader vision for in-app AI, with plans for future enhancements and integrations.

Google's Tensor ML SDK has transitioned from its Experimental Access Program to Beta, allowing developers to create and deploy AI experiences on the Google Pixel 10 family using the Tensor Processing Unit (TPU) on the custom-designed Google Tensor System-on-Chip. The SDK, integrated with Google's LiteRT framework, facilitates a seamless workflow by abstracting low-level SDKs and offering a unified API for machine learning model deployment. Developers can now convert, compile, and run PyTorch or TFLite models on the TPU, with fallback options for CPU or GPU, and distribute applications using Play Feature Delivery and AI Packs. The SDK's Model Garden offers over 100 precompiled models, including those for computer vision, speech recognition, and generative AI, enabling developers to craft AI applications with features like real-time text generation, image filters, object detection, and voice-controlled tools. The initiative encourages the developer community to experiment with these capabilities, providing resources such as documentation, codelabs, and community support on platforms like GitHub.

Gemini CLI, initially launched to bring the capabilities of Gemini to terminal users, is evolving into a more integrated platform called Google Antigravity, designed to address the needs of a multi-agent environment. Antigravity CLI, built in Go for enhanced performance, retains key features of Gemini CLI while introducing asynchronous workflows and a unified architecture that aligns with the new Antigravity desktop application. As of June 18, 2026, Gemini CLI and its associated extensions will cease operation for certain users, pushing them to transition to Antigravity CLI, which is already available and supported by comprehensive documentation and upcoming video guides. Enterprise customers will maintain access to Gemini CLI with ongoing support, but are encouraged to explore the new capabilities of Antigravity CLI, especially as it offers a seamless integration with Google Cloud projects and promises continual updates to its core agents.

Google AI Edge Gallery has introduced new features such as support for the Model Context Protocol (MCP), notifications reminders, and persistent chat history, enhancing the capability of on-device AI interactions. These updates allow developers and users to create more connected and automated experiences with models like Gemma 4 on Android and iOS, providing a platform for agentic workflows directly on mobile devices. The integration of MCP enables standardized interactions with external tools, facilitating complex tasks across various data sources. The app now supports scheduling routines with notification reminders, allowing proactive use cases such as mood tracking and calendar briefings. Additionally, enhancements like persistent chat history and customizable system prompts offer greater flexibility for experimenting with on-device models. The open-source nature of the app encourages community-driven innovation, with developers contributing to a range of utility-focused workflows and skills.

Google AI Edge's LiteRT-LM offers a highly optimized experience for deploying the Gemma 4 model across platforms, leveraging the LiteRT framework for inference. This engine supports advanced AI functionalities in various Google products, such as Chrome, ChromeOS, and Pixel Watch, as well as the Google AI Edge Gallery app. LiteRT-LM enhances performance through features like Multi-Token Prediction (MTP) and advanced session management, enabling efficient memory utilization and high-speed decoding across CPU, GPU, and NPU backends. The platform supports complex task execution with Thinking Mode and constrained decoding, and is designed for cross-platform development with APIs for Android, iOS, and web applications. With its comprehensive integration, LiteRT-LM promises to advance the development of privacy-focused, low-latency applications on edge devices.

At Google I/O 2026, the Google Cloud and NVIDIA developer community celebrated reaching 100,000 members and highlighted resources to support developers in optimizing large language models (LLMs), deploying scalable workloads, and exploring agentic AI. The community offers four curated learning pathways to help developers transition from concept to production quickly, covering topics such as deploying generative AI models with NVIDIA NIM on GKE, accelerating machine learning, and optimizing data analytics on GPUs. Monthly livestream sessions with technical experts provide insights into real-world engineering challenges, while community forums facilitate discussions and knowledge sharing among peers. The initiative underscores the collaboration between Google Cloud and NVIDIA to advance the development of generative AI, with future plans including advanced technical content, hands-on labs, and quarterly events offering direct access to engineering experts.

Google's I/O event introduced significant advancements in AI and development platforms, including the Gemini 3.5 series of models and the upgraded Antigravity platform, which enhances the orchestration and building of autonomous agents for complex workflows. The event highlighted new tools for developing Android and web apps, such as Antigravity 2.0 with a CLI and integrations with Google AI Studio, allowing for seamless app development and deployment using AI agents. Google also unveiled Android-specific tools like Android CLI and Android Bench, an LLM leaderboard for development tasks, alongside a new migration agent in Android Studio for faster app code transitions. In the realm of web development, the proposed WebMCP standard was introduced to enable browser-based AI agents to perform complex tasks efficiently, complemented by Modern Web Guidance to improve web experience performance and security. Additional innovations such as Chrome DevTools for agents and the HTML-in-Canvas API were also showcased, enhancing the capabilities of AI agents in real-time code optimization and the creation of interactive 3D experiences.

AI technology is advancing towards multimodal capabilities that include on-device image and audio generation, allowing developers to create personalized consumer experiences. Traditionally, executing large AI models at the edge has involved a tradeoff between high latency on CPUs and using specialized, fragmented accelerators. The Arm Scalable Matrix Extension 2 (SME2) resolves this by integrating matrix-compute units into the CPU, enhancing its performance as an AI accelerator and improving inference speeds for generative AI tasks by up to 5x. Google's AI Edge platform further simplifies AI deployment on Arm hardware, supporting automatic runtime optimizations through tools like LiteRT, XNNPACK, and Arm KleidiAI, which enhance efficiency by targeting math-intensive kernels. By leveraging this integration, developers can transform models like Stability AI's stable-audio-open-small into optimized, mixed-precision implementations suitable for high-performance edge deployment, while Google's AI Edge Quantizer and Model Explorer facilitate model compression and performance optimization. This synergy enables significant performance improvements, reducing audio generation time and memory usage while maintaining audio quality, opening opportunities for scaling applications across a wide range of CPU-powered devices globally.

Genkit is an open-source framework designed for creating full-stack, AI-powered applications across various platforms, supporting languages such as TypeScript, Go, Dart, and Python. It emphasizes the necessity of incorporating features like retries, fallbacks, human approvals, and observability to ensure reliability and manage AI-driven applications effectively. To address these needs, Genkit utilizes a middleware system that allows developers to inject custom behaviors into generation calls through composable hooks at different stages of the tool loop. The middleware is currently available in TypeScript, Go, and Dart, with Python support forthcoming. Pre-built middleware solutions include functionalities for retrying failed API calls, switching models on errors, approving tool execution, managing skill sets, and controlling filesystem access. Developers can also create custom middleware to address specific requirements, such as filtering unwanted content, by defining hooks that modify or validate model outputs. Genkit encourages sharing custom middleware as packages for community use and offers a Developer UI for inspecting and debugging middleware configurations and executions.

The tutorial discusses building a New Hire Onboarding Coordinator Agent using the Agent Development Kit (ADK) to address the limitations of stateless chatbots in handling long-term enterprise workflows. Stateless chatbots are inadequate for processes with extended idle times, such as HR onboarding or invoice dispute resolutions, because they rely on appending conversation history, which leads to context pollution, token cost issues, and reasoning errors. The tutorial introduces three architectural shifts: durable memory schemas, event-driven dormancy gates, and multi-agent delegation. By implementing a state machine to track onboarding progress and using persistent session storage, the agent can maintain context across pauses and resumes, ensuring reliable workflow management. This approach includes multi-agent coordination, allowing specialized sub-agents to handle tasks like IT provisioning, which enhances focus and reasoning quality. The tutorial also covers using webhook endpoints to handle idle time effectively and evaluates multi-day flows through tests that simulate delays, ensuring the agent remains functional over extended periods. The tutorial concludes by encouraging users to explore the ADK documentation and deploy their long-running agents for various workflows requiring persistence and state management.

Researchers at UCSD, led by Hao Zhang, have made a significant advancement in the field of Large Language Model (LLM) acceleration by implementing a novel speculative decoding method called DFlash on Google TPUs. This method, which utilizes block diffusion, allows for the generation of an entire block of candidate tokens in a single forward pass, overcoming the traditional bottleneck of sequential token prediction. By integrating DFlash into the vLLM TPU inference ecosystem, the UCSD team achieved a remarkable average speed increase of 3.13x in tokens per second on TPU v5p, with peak speedups reaching nearly 6x for complex math tasks. This development not only showcases the potential of diffusion-style drafting to leverage the parallel computing capabilities of TPUs but also sets the stage for future innovations in speculative decoding systems, paving the way for Speculative Speculative Decoding (SSD) and broader applications in AI hardware acceleration.