April 2026 Summaries

Gemini Embedding 2, now generally available through the Gemini API and Gemini Enterprise Agent Platform, is a versatile embedding model that maps various data types—including text, images, video, audio, and documents—into a unified embedding space across over 100 languages. This model handles diverse inputs in a single call and enables developers to create applications that can "see" and "hear" complex, real-world data, offering improved accuracy in tasks such as agentic multimodal retrieval-augmented generation (RAG) and visual search. The model's ability to process interleaved inputs enhances its understanding of data, which can be leveraged in AI-driven tasks like multimodal search, search reranking, and anomaly detection. Users like Harvey and Supermemory have already reported increased accuracy in legal research and conceptual searching, respectively, after integrating Gemini Embedding 2. Additionally, the model supports efficient storage solutions by allowing dimensionality reduction of its vectors, thus reducing costs without compromising accuracy. As industries explore the model's potential, Gemini Embedding 2 promises to enhance the understanding and processing of complex data across various applications.

Google Cloud has announced a significant advancement for AI/ML workloads using the PyTorch ecosystem by integrating Rapid Storage, powered by Google's Colossus storage architecture, via the fsspec interface. This integration aims to address the bottleneck challenges of data loading and checkpointing that arise as model sizes increase, ensuring GPUs remain efficiently utilized. The new Rapid Bucket solution offers high-performance object storage using gRPC bidirectional streams, bypassing traditional REST APIs, which significantly enhances throughput and reduces latency. With its direct connectivity and zonal co-location, Rapid Storage achieves an aggregate throughput of over 15 TiB/s and ultra-low latency of under 1ms for various operations. This is seamlessly integrated into existing systems without requiring extensive code rewrites, allowing developers to enjoy significant performance improvements by simply switching to Rapid Buckets. Testing has demonstrated a 23% performance gain compared to standard regional buckets, with notable improvements in both read and write throughput.

LiteRT is a cross-platform framework designed to enhance on-device AI performance by leveraging Neural Processing Units (NPUs) across various platforms such as mobile, desktop, and IoT. It allows developers to integrate advanced AI models without the need for vendor-specific code, providing both CPU, GPU, and NPU acceleration. This framework is utilized by industry leaders like Google Meet, Epic Games, and Argmax Inc. for applications ranging from real-time video effects and facial animation to speech recognition, demonstrating significant improvements in efficiency and responsiveness. LiteRT simplifies the deployment of AI features by abstracting complex NPU integrations and supporting a wide range of hardware, thus enabling developers to optimize performance across different devices. The Google AI Edge Gallery App and Google AI Edge Portal offer developers tools to test, validate, and benchmark their AI models, ensuring optimal performance across various configurations.

AI agents are evolving from experimental scripts to production-ready services, yet the infrastructure supporting their development remains fragmented, causing inefficiencies for developers and coding assistants. Google Cloud's newly introduced Agents CLI within the Agent Platform aims to streamline the Agent Development Lifecycle (ADLC) by providing a cohesive programmatic framework for AI coding agents like Gemini CLI and Claude Code. This tool facilitates the seamless integration of cloud components, allowing developers to efficiently scaffold, simulate, and evaluate AI projects with machine-readable inputs and API references. By using Agents CLI, developers can rapidly transition from local prototypes to globally distributed services through automated deployment processes, including Infrastructure as Code (IaC) and CI/CD pipelines. The CLI supports both agent-driven and human-driven modes, enabling developers to maintain control while benefiting from the efficiency of AI. Users are encouraged to explore the platform's documentation and community resources to harness the potential of their AI assistants in building the next generation of production-grade AI systems.

Building AI agents that function effectively in real-world applications requires more than just elegant coding; it involves addressing challenges such as rate limits, scaling, and avoiding operational failures. The AI Agent Clinic was launched to tackle these challenges, with the first episode focusing on a sales research agent named "Titanium." Originally a monolithic Python script limited to hardcoded data, Titanium was restructured into a distributed framework using Google’s Agent Development Kit (ADK), enhancing reliability and scalability. The transformation included creating specialized sub-agents, implementing structured outputs with Pydantic, and replacing hardcoded information with a dynamic data intake system. Observability was improved through OpenTelemetry, and cost optimization was achieved by leveraging ADK's orchestration features. The series aims to help others diagnose and refactor problematic agents by inviting submissions for live analysis and improvement.

Generative UI technology, specifically A2UI v0.9, enables AI agents to dynamically create customized user interface widgets using an existing component catalog without altering the design system. A2UI v0.9 offers a framework-agnostic standard for UI intent declaration, allowing seamless communication between agents and client applications across various platforms, including web and mobile. This release emphasizes ease of integration with existing frontends, introduces a shared web-core library for simplified UI rendering, and enhances developer experience with improved abstractions and streamlined streaming. The A2UI Agent SDK facilitates optimized UI generation with low latency, featuring new language capabilities for validation and collaborative editing. The ecosystem continues to expand with contributions from companies like Vercel and Oracle and applications such as the GenUI Personal Health Companion and Life Goal Simulator, both demonstrating real-time, context-aware UI generation. A2UI's compatibility with AG-UI middleware enables any agent to leverage these capabilities, promoting flexible and portable generative UIs.

In the evolving field of large language models (LLMs), post-training techniques are crucial to enhance pre-trained models into specialized assistants or reasoning engines. MaxText introduces new post-training features, including Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), designed for single-host TPU configurations like v5p-8 and v6e-8, utilizing the JAX library and Tunix for efficiency. SFT allows users to fine-tune models with labeled datasets using seamless integration with Hugging Face datasets and flexible checkpoints, while RL supports advanced reasoning capabilities with algorithms such as Group Relative Policy Optimization (GRPO) and Group Sequence Policy Optimization (GSPO), optimizing training stability and efficiency. These advancements offer a scalable, high-performance path for developers to refine their models, with the potential for transitioning to multi-host configurations for larger models and datasets in the future.

Gemini CLI utilizes subagents, which are specialized expert agents designed to handle complex, repetitive, or high-volume tasks, allowing the primary agent to remain focused on overarching goals and decision-making. Each subagent operates within its own context, using custom system instructions and tools to efficiently execute tasks and return a consolidated response to the main session. Users can create custom subagents for specific workflows, defined through Markdown files with YAML frontmatter, which can be shared or bundled within Gemini CLI extensions. Subagents can run in parallel, drastically speeding up processes by managing multiple tasks simultaneously, although caution is advised to avoid conflicts during parallel code edits. Built-in subagents, such as the generalist, cli_help, and codebase_investigator, are available to assist with various tasks, and users can explicitly delegate tasks to subagents using the @agent syntax. This system ensures efficient task management and supports the creation of tailored solutions to meet project-specific needs.

Google has announced enhancements to the Google Pay API focusing on merchant initiated transactions (MIT), offering developers increased flexibility and control in creating seamless payment experiences for use cases like subscriptions, deferred payments, and automatic reloads. While the API previously supported MIT, these updates allow for clearer specification of future payment terms, ensuring transparency and smoother processing. New objects such as recurringTransactionInfo, deferredTransactionInfo, and automaticReloadTransactionInfo have been introduced to detail recurring payment schedules, future one-time payments, and balance-based reloads, respectively. These changes aim to improve payment success rates, provide improved transparency for users, and support a wider variety of billing models beyond immediate payments, while also ensuring compliance with future network mandates. Developers are encouraged to explore the updated API documentation to leverage these features for building more robust and user-friendly payment solutions.

Google I/O is set to take place on May 19–20, featuring key updates across AI, Android, Chrome, and Cloud, with a focus on introducing agentic coding tools to enhance software development. The event will kick off with a Google keynote followed by a Developer keynote, offering two days of live sessions from Mountain View. Attendees will explore the agentic era of development, the integration of AI into app workflows, and new tools for creating advanced web applications. The agenda includes sessions on the latest developments in Android, Google AI, and Chrome, with additional on-demand content and codelabs available from May 21.

The Google Cloud AI Agent Bake-Off challenged developers to create fully autonomous AI agents under tight deadlines to solve complex industry problems such as e-commerce returns, banking system modernization, and startup strategy automation. This experience highlighted the importance of moving beyond simple interactions with large language models (LLMs) to employing rigorous agentic engineering, which involves multi-agent architecture, state management, and deterministic guardrails. Successful projects utilized a modular approach, treating agents like microservices to break down complex tasks into specialized sub-agents, which improved efficiency and maintenance. Developers were encouraged to prepare for rapid advancements in AI by embracing impermanence and building systems that can easily adapt and integrate multimodal capabilities for a more organic user experience. Additionally, adopting open-source protocols and strict validation schemas ensures robust and scalable production systems, as these practices allow AI to handle reasoning while deterministic code executes precise tasks. The Bake-Off emphasized that building reliable, secure, and useful AI agents requires a commitment to software architecture fundamentals rather than relying on complex prompts or single-shot demos.

TorchTPU is an advanced integration designed to enable PyTorch to run natively and efficiently on Google's Tensor Processing Units (TPUs), which are critical to handling modern AI infrastructure demands involving large-scale distributed systems. Developed with a focus on usability, portability, and performance, TorchTPU allows developers to migrate existing PyTorch workloads with minimal code changes while maximizing the computational capabilities of TPUs. The system offers three eager execution modes (Debug, Strict, and Fused Eager) to enhance flexibility and performance, with Fused Eager providing significant performance improvements through automated operation fusion. TorchTPU also integrates with PyTorch's torch.compile interface for full-graph compilation, utilizing the XLA backend to optimize dense computations and communications. This integration supports various distributed training setups and overcomes previous limitations by accommodating divergent executions (MPMD) alongside standard SPMD optimizations. TorchTPU's architecture encourages optimal model designs tailored to TPU hardware characteristics, and future developments aim to reduce recompilation overhead, expand custom kernel capabilities, and support dynamic shapes directly through PyTorch's interfaces. As Google works towards a comprehensive PyTorch experience on TPUs, the team is striving to address open challenges and expand the system's capabilities by 2026 and beyond, enhancing AI model efficiency and scalability.

In March, Google announced that US users can now update their Google Account usernames while retaining their account, inbox, and data intact, with both the old and new email addresses remaining active and pointing to the same inbox. This update, previously available to business Workspace users, is now extended to consumer Google Accounts with an @gmail.com address. For platforms using Sign in with Google, the impact varies based on user authentication methods; those using subject ID as the primary identifier need no changes, while platforms using email addresses might face issues like duplicate accounts. To ensure continuity, Google recommends adopting subject ID as the primary user identifier, enabling email updates on platforms, and offering email-based account recovery. These measures aim to enhance user experience by providing a seamless transition and maintaining access continuity.

Google DeepMind has introduced Gemma 4, a family of advanced open-source AI models designed for on-device applications, under the Apache 2.0 license. These models support a wide range of functionalities, including multi-step planning, autonomous actions, offline code generation, and audio-visual processing, without needing specialized fine-tuning. Gemma 4 supports over 140 languages and can be used across mobile, desktop, and edge devices, with integration into platforms like Android AICore and Google AI Edge Gallery. The LiteRT-LM framework enhances the model's performance on various devices, such as Raspberry Pi and Qualcomm processors, by optimizing memory use and processing speeds. Developers can create agentic workflows using tools like Agent Skills, which enable enriching user experiences through interactive content, knowledge augmentation, and seamless integration with other models. Gemma 4's capabilities are accessible through a new Python package and CLI tool, allowing developers to experiment and deploy AI-driven applications with ease.

The Agent Development Kit (ADK)'s SkillToolset enables AI agents to dynamically generate and load domain expertise, allowing them to create new instructions and capabilities as needed. By utilizing a progressive disclosure architecture, agents can load context-specific information only when necessary, avoiding the inefficiency of monolithic system prompts that cram all instructions into a single string. This method involves three levels of knowledge loading: L1 Metadata for a quick overview of all skills, L2 Instructions for complete skill details, and L3 Resources for external references. The guide outlines four skill patterns: inline skills for simple, stable rules; file-based skills for complex tasks that need additional references; external skills sourced from community repositories; and a skill factory that allows agents to generate new skills independently. This architecture reduces token usage while enabling agents to expand their capabilities dynamically, making them more efficient and adaptable for tasks such as security reviews and compliance audits. The SkillToolset auto-generates tools that map directly to these levels, promoting efficient and flexible AI agent development.