June 2026 Summaries

Building and maintaining high-quality software agents requires a disciplined approach that bridges the gap between anecdotal success and consistent performance in production. This methodology, discussed at Cloud Next '26, is encapsulated in a three-phase flywheel: Build & Test, Ship & Monitor, and Learn & Refine, and further enhanced by a developer-facing path known as the quality-flywheel skill. This skill integrates automated evaluation processes with Google's AutoRaters in collaboration with Google DeepMind, allowing for continuous improvement of agents by conducting targeted testing, analyzing failures, and proposing optimizations without human-in-the-loop grading. The system is designed to identify and rectify subtle failures that might not be immediately obvious, such as discrepancies between an agent's internal state and its output to users, by using custom rubrics and synthetic scenarios to simulate user interactions. As agents mature, the focus shifts from simulated to real production data to ensure that each user interaction serves as a benchmark for further refinement. The quality-flywheel skill is adaptable, serving both specific goals and broader diagnostic purposes, ultimately aiming to create an environment where agents are continuously improvable rather than perfect.

ADK for Go 2.0 introduces a graph-based workflow engine that enhances the development of production agent applications by allowing developers to compose multi-agent workflows in a structured, reliable, and observable manner. This new version incorporates features such as human-in-the-loop (HITL) capabilities, dynamic orchestration in plain Go, and a unified node runtime, aligning it closely with the Python ADK's graph-first approach. The graph structure enables real-world applications to handle complex tasks like branching, looping, human approvals, and retries efficiently, while the use of nodes and edges simplifies the representation and execution of workflows. ADK 2.0 also introduces new agent modes and a streamlined programming model that consolidates contexts, making it easier to manage tools, callbacks, and graph nodes. This upgrade from ADK 1.0 is designed to be additive, offering richer event streams and improved resilience features, allowing developers to build robust and scalable agent applications in Go.

In the context of a contract compliance pipeline, a Python agent and a Go agent collaborate effectively using the Agent2Agent (A2A) protocol, allowing them to function together without rewriting their respective codebases. The Python agent extracts terms from legal contracts using Gemini, while the Go agent validates these terms against corporate policies. The integration is facilitated by Google’s Agent Development Kit (ADK), which simplifies the orchestration of multi-agent systems through cross-language collaboration. This setup addresses common issues of monolithic AI systems, such as context degradation, unhandled exceptions, and testing challenges, by decomposing them into specialized microservices. The A2A protocol serves as a universal communication standard, similar to HTTP, enabling agents to interact regardless of their internal implementations. This approach not only maintains the technological independence of different teams but also enhances resilience and auditing capabilities, with the entire system being available as an open-source application for deployment and testing.

AI coding agents are evolving from reactive assistants to proactive engines that not only complete tasks but also anticipate developers' needs by continuously absorbing context and identifying emerging risks. This shift focuses on moving from task-based to goal-oriented frameworks, where agents explore codebases to provide diagnostic insights that guide developers toward higher-level objectives. Current benchmarks, like SWE-Bench, evaluate task completion but do not assess goal achievement, prompting researchers to emphasize the importance of an agent's insight policy, which determines when to provide feedback to developers. Research conducted by Google Labs, involving the analysis of 705 bugs from internal codebases, aims to establish a "ground truth" by clustering related bugs to reveal common goals, thus allowing for the evaluation of proactive agents. Preliminary results indicate that the diagnostic logic is effective, with agents scoring high relevance in their insights, especially when given more exploration rounds. The research is expanding to include public GitHub data and richer context streams, aiming to make this methodology applicable to a broader AI community.

The rise of the Agent-to-Agent (A2A) protocol marks a significant shift in the development of AI agents, moving beyond the limitations of stateless tools to create a dynamic, collaborative agentic ecosystem. Launched a year ago, A2A addresses the need for a flexible, secure, and autonomous communication method among AI agents, overcoming the constraints of traditional APIs. It introduces a framework that maintains data privacy, minimizes context pollution, and supports dynamic autonomy, enabling agents to collaborate effectively and handle complex tasks without overloading their primary functions. The A2A protocol is particularly exemplified in applications like FoldRun, which simplifies the process of predicting protein structures in life sciences by managing intricate dependencies and decision-making autonomously. This approach extends to various sectors, facilitating innovations in agentic commerce, enterprise data management, cross-platform IT operations, and secure telecommunications, all while prioritizing security and interoperability. As the A2A ecosystem grows, developers are encouraged to engage with the platform through available SDKs, contributing to the evolution and expansion of this collaborative AI landscape.

Agentic Resource Discovery (ARD) is an open specification designed to streamline the discovery, verification, and connection of AI capabilities across the web, addressing the current fragmentation and lack of interoperability in agentic systems. By standardizing the way capabilities are published and indexed through catalogs and registries, ARD facilitates secure and efficient communication between agents across different organizations and platforms. This specification enables agents to dynamically find and verify resources by relying on cryptographic metadata for trust, allowing for direct and secure connections using native protocols. Google's Gemini Enterprise Agent Platform supports ARD through its Agent Registry, providing enterprises with tools for governance and compliance while fostering an open and decentralized agent ecosystem. ARD is available for use, with opportunities for the community to contribute to its ongoing development.

Developers often face a trade-off between customization and integration when evolving agentic workflows, traditionally having to choose between Model Context Protocol (MCP) Apps, which offer creative freedom but can lead to design inconsistency and performance issues, and Agent-to-User Interface (A2UI), which provides a secure and consistent UI but restricts developers to specific component libraries. By integrating these two approaches, developers can leverage native component rendering for standard UI elements while reserving custom iframe embedding for tailored experiences, enhancing both security and performance. Three architectural patterns are proposed to facilitate this integration: A2UI over MCP servers allows native rendering of rich UIs without the need for complex Agent-to-Agent architecture, MCP Apps within A2UI components provide a creative yet secure environment for complex modules, and A2UI inside MCP Apps enables dynamic, agent-driven UIs in legacy applications. These patterns allow developers to bypass iframe limitations, maintain robust security, and accelerate development cycles, all while ensuring design consistency and flexibility in UI presentation.

In response to the increasing threats of phishing and online abuse, Google has introduced new session metadata claims, auth_time and amr, within its Sign in with Google feature, aimed at enhancing security for verified apps using OpenID Connect (OIDC) standards. These claims, included in the ID Token, provide insights into when and how users authenticate, thereby enabling platforms to implement more dynamic, risk-based access controls. Auth_time offers a timestamp of the last successful authentication, helping platforms enforce session policies, while amr details the authentication methods used, such as passwords or multi-factor authentication, facilitating more granular access controls. These enhancements aim to reduce security incidents and fraudulent activities by providing platforms with better tools to manage and secure user sessions across Android, iOS, and web applications. By leveraging Google's authentication infrastructure, applications can benefit from reduced security burdens and improve their overall security posture without drastically altering existing authentication flows.

Google has launched the TPU Developer Hub, an educational resource aimed at enhancing the capabilities of model builders, optimizers, and developers working with Google Cloud TPUs. This hub serves as a centralized platform offering up-to-date guidance and tools necessary to optimize TPU infrastructure and its software stack. It provides a comprehensive range of resources covering the entire developer lifecycle, from pre-training to inference workloads, and includes topics such as hardware architecture, software stack capabilities, tracing, debugging, optimization strategies, and networking security. The hub is designed to facilitate practical, code-first learning with open-source code recipes and technical documentation that can be integrated into workflows, supporting both manual browsing and AI-assisted development tools. By making TPUs more accessible through an open ecosystem, the TPU Developer Hub aims to empower developers to achieve peak model performance and efficiency while maintaining robust security and networking foundations.

DiffusionGemma is an experimental model built on the Gemma 4 backbone designed to enhance developer workflows by shifting the bottleneck from memory bandwidth to compute, allowing for up to 4x faster token generation on GPUs. Utilizing a 26B Mixture of Experts model, it activates only 3.8B parameters during inference, making it deployable within an 18 GB VRAM limit. DiffusionGemma features bidirectional context and self-correction, enabling real-time error correction and parallel context propagation. Its Uniform State Diffusion approach refines a 256-token canvas in parallel, and for sequences longer than 256 tokens, it employs block autoregressive diffusion. This architecture is particularly effective for multivariable constrained problems like Sudoku, as it allows for global context awareness and self-correction. The model's integration with vLLM enables efficient deployment and iterative parallel denoising loops across batched request streams. Fine-tuning on Sudoku puzzles has shown an 80% success rate, demonstrating the model's capability in handling non-sequential tasks efficiently. The model is optimized for deployment across various hardware, from consumer-grade graphics cards to enterprise servers.

Google has announced the Google Colab Command-Line Interface (CLI), which connects local terminals with remote Colab runtimes to streamline execution for developers and AI agents. The Colab CLI facilitates instant GPU or TPU access, remote execution of Python scripts, easy retrieval of models and datasets, and interactive remote environments. It integrates smoothly with standard terminal environments and includes a prepackaged Colab skill file for AI agents, enabling them to efficiently utilize the CLI. A practical example involves using the CLI to fine-tune the Gemma 3-1B model on a Text-to-SQL dataset, showcasing the capability to offload computational tasks to a GPU with minimal commands. This tool aims to make Colab's computational power more accessible and is available via the Google Colab CLI GitHub repository.

Google DeepMind's latest model, Gemma 4 12B, is an advanced open model designed to bring multimodal intelligence to everyday laptops through the Google AI Edge stack. This model enables a range of powerful on-device capabilities, from autonomous data processing and visual insight generation to script execution for data analysis and the creation of fully functional webpages. The Google AI Edge Gallery app, available on macOS, showcases Gemma's coding capabilities, allowing users to interact with data through natural language prompts. Additionally, the Google AI Edge Eloquent app provides a fully offline dictation and editing experience, transforming spoken commands into polished text, demonstrating the model's enhanced instruction-following and quality improvements over previous iterations. The LiteRT-LM CLI tool further extends these capabilities by allowing users to run language models locally, creating industry-compatible endpoints that integrate with various tools. Together, these innovations enable users to leverage the full potential of AI-driven applications directly on their devices while maintaining data privacy and cost efficiency.

Gemma 4 12B is a newly launched dense multimodal AI model with a unique encoder-free architecture, designed to enhance local AI functionalities by directly integrating multimodal data into its LLM backbone, thereby reducing latency. It marks a significant advancement in the Gemma family as the first medium-sized model capable of ingesting audio inputs natively, and it is optimized for local use on devices equipped with 16GB of VRAM. The model supports various capabilities, including automatic speech recognition, agentic reasoning, and video understanding, demonstrating its versatility in multimodal applications. Additionally, it offers a new MacOS desktop experience, allowing developers to run AI tasks locally on consumer-grade devices, and introduces the LiteRT-LM for zero-latency local execution. The model's unified architecture enables seamless tuning across vision, audio, and text inputs, offering developers the flexibility to build local multimodal agents using tools like Hugging Face and llama.cpp and deploy them through platforms such as Google Cloud and the Gemini Enterprise Agent Platform.