November 2025 Summaries

Gemini 3, the latest and most advanced model from the Gemini suite, offers developers enhanced capabilities through the Gemini API, providing state-of-the-art reasoning, autonomous coding, and multimodal understanding. Key updates include a new thinking_level parameter that allows developers to control the depth of the model's reasoning, enabling them to optimize for either complexity or efficiency depending on the task. The media_resolution parameter provides granular control over how images, videos, and documents are processed, balancing visual fidelity with token usage. Thought Signatures, encrypted representations of the model's internal thought process, are introduced to maintain reasoning chains across interactions, crucial for complex workflows. Additionally, the API supports grounding with Google Search, which now features a more flexible pricing model, and allows integration of live web data into structured outputs. These enhancements, along with recommended best practices like maintaining consistency and clarity in multi-modal inputs, aim to optimize the performance of Gemini 3 in diverse applications, from vibe coding to complex problem solving.

Modern smartphones are equipped with sophisticated SoCs that include CPUs, GPUs, and NPUs, enabling advanced on-device GenAI experiences that surpass server-only counterparts in interactivity and real-time performance. While GPUs are prevalent in Android devices for accelerating AI tasks, they can become bottlenecks when handling complex applications like text-to-image generation alongside live camera feed processing. The introduction of NPUs, which are highly specialized for AI tasks, offers a solution by providing significantly more power-efficient AI compute compared to CPUs and GPUs. This enhanced architecture allows for concurrent processing, freeing GPUs for rendering and CPUs for main-thread logic, facilitating smoother and faster AI application performance. Google's LiteRT Qualcomm AI Engine Direct Accelerator further advances this capability by integrating NPU power and simplifying mobile deployment workflows, allowing developers to deploy models seamlessly across different SoCs. This accelerator supports extensive LiteRT operations and specialized kernels, providing substantial performance gains—up to 100 times faster than CPUs and 10 times faster than GPUs—across various ML models, thus enabling previously unreachable real-time AI experiences on mobile devices.

Google Antigravity is a newly introduced agentic development platform that elevates software development by allowing users to operate at a higher, task-oriented level. Unlike traditional coding tools, it integrates an AI-powered coding experience with an agent-first interface, enabling autonomous agents to plan, execute, and verify complex tasks across the editor, terminal, and browser. The platform features two main interfaces: the Editor View for hands-on coding and the Manager Surface for orchestrating multiple agents asynchronously. Antigravity aims to reduce context switching by allowing developers to delegate end-to-end tasks, such as UI changes and long-running maintenance tasks, to agents that provide Artifacts—tangible deliverables like screenshots and task lists—for easy verification. Emphasizing learning and adaptability, Antigravity supports multiple operating systems and AI models, and is available for public preview at no cost for individuals, marking a significant shift in development workflows.

JAX has emerged as a vital framework for developing advanced foundation models in AI, with prominent companies like Anthropic, xAI, and Apple utilizing it for their machine learning efforts. The JAX AI Stack, an end-to-end platform based on JAX, enhances this framework by offering a modular and flexible architecture that enables users to create custom machine learning stacks. Key components include JAX for array computation, Flax for neural network authoring, Optax for optimization, and Orbax for checkpointing, all integrating seamlessly with industrial-scale infrastructure to support large-scale distributed computation. The stack also includes advanced tools like Pallas and Tokamax for kernel customization, Qwix for model quantization, and Grain for efficient data loading, facilitating the entire ML lifecycle from research to deployment. Examples of its success include increased throughput for Kakao's LLMs and enhanced scalability for Lightricks' video model, showcasing its capability to optimize cost-performance and drive innovation across AI applications.

Gemini 3 Pro Preview, a powerful agentic AI model, has been introduced as an advanced tool for creating sophisticated, semi-autonomous systems capable of complex decision-making. Designed to offer developers precise control over cost, latency, and reasoning depth, the model features adjustable reasoning levels, stateful tool use via encrypted "Thought Signatures," and adjustable multimodal fidelity for diverse media analysis. With a large context window to maintain logic consistency, Gemini 3 integrates with an open-source ecosystem, ensuring immediate compatibility with frameworks like LangChain, AI SDK by Vercel, LlamaIndex, Pydantic AI, and n8n. These integrations enable developers to construct reliable, stateful AI agents efficiently, while also allowing non-developers to create sophisticated agents without coding. The model supports best practices such as simplifying prompts, maintaining a reasoning temperature of 1.0, and optimizing visual tokens, all detailed in the comprehensive Gemini 3 Developer Guide.

Google has introduced Gemini 3, its most advanced model, which will be available in Jules, an autonomous coding agent for Google AI subscribers, starting with Ultra subscribers and soon for Pro plan users. Gemini 3 Pro enhances Jules by improving agentic capabilities, making it more reliable and coherent in handling multi-step tasks while maintaining context with less user effort. Jules is designed to be always available across various platforms, such as terminals and CLI extensions, allowing seamless transitions without losing project context. Recent updates to Jules, including parallel runs, Windows support, and a more stable API, enhance predictability and transparency, while features like enhanced Git handling and improved VM performance aim to streamline user experience. Future enhancements include automatic PR creation and simplified start modes, underscoring the commitment to making Jules a more intuitive tool for developers.

Gemini 3 Pro, the latest model from Gemini, is now integrated into Gemini CLI, offering enhanced performance and productivity for terminal users, particularly those subscribed to Google AI Ultra or holding a paid Gemini API key. The integration allows for state-of-the-art reasoning in command execution, agentic coding for complex engineering tasks, and smarter workflows through advanced tool use. Key features include the ability to synthesize information from various sources, execute complex coding tasks, generate detailed applications like a 3D simulation of the Golden Gate Bridge, and translate natural language into shell commands. Additionally, Gemini 3 Pro can read and document codebases, orchestrate complex workflows for debugging, and improve everyday operations such as refactoring and error management, making it a versatile tool for developers. Users can begin experiencing these capabilities by upgrading their Gemini CLI version and enabling preview features, with further access rolling out gradually and a waitlist available for other users.

The Gemini CLI has undergone a significant upgrade to enhance the user experience by making terminal interactions more robust, intuitive, and visually stable. This includes a revamped rendering foundation that eliminates common terminal application issues like screen flickering and bouncing input prompts, and introduces features such as mouse-based navigation, sticky headers, and a stable input prompt. These updates are designed to provide a smoother and more graphical-interface-like experience within the terminal, with improvements like flicker-free displays and robust rendering that adapts to terminal resizing without glitches. Despite using an alternate screen buffer for this improved experience, users can still access their complete chat history after exiting the Gemini CLI. The new features are available by default starting from version 0.15.0, with further enhancements like click-and-drag functionality and improved copy-paste solutions in development.

Google has launched a new extension for Visual Studio Code (VS Code) that integrates Google Colab, a popular platform for AI/ML development, with one of the world's most widely used code editors. This new extension allows users to leverage the strengths of both VS Code and Colab, offering a seamless experience for coding, collaboration, and accessing powerful compute resources like GPUs and TPUs. Until now, developers often used separate environments for project development in VS Code and for executing and visualizing notebook workloads in Colab. The extension bridges this gap by enabling local notebooks in VS Code to connect with Colab runtimes, which is particularly beneficial for those who require enhanced IDE features and extensibility. This integration aims to provide a unified workflow for developers, students, and researchers who are already familiar with VS Code, while still granting access to Colab's computational resources. The project is positioned as a starting point for further integration, with plans to enhance the Colab experience within VS Code in the future.

Code Wiki is an innovative platform launched by Google to mitigate the challenges associated with reading and understanding existing code, a significant bottleneck in software development. It provides a continuously updated, structured wiki for code repositories, allowing for interactive navigation and integration with an intelligent chat agent powered by Gemini. This system automates the documentation process by regenerating it after each code change, ensuring that the documentation evolves with the code. It offers direct hyperlinks to relevant code files, classes, and functions, and generates up-to-date architecture diagrams to visualize complex relationships. The platform aims to make the process of understanding code seamless and efficient, enabling new contributors to make their first commit quickly and allowing senior developers to grasp new libraries swiftly. Additionally, a forthcoming Gemini CLI extension will allow organizations to deploy the same system securely on internal repositories, facilitating a deeper understanding of legacy code.

Metrax is a high-performance library developed to provide efficient and robust model evaluation metrics for JAX, addressing the lack of a built-in metrics library as teams transitioned from TensorFlow. It offers predefined metrics for various types of machine learning models, ensuring compatibility and consistency in distributed and scaled training environments, which allows users to focus on evaluation results rather than implementing metric definitions. Metrax integrates well with the JAX AI Stack and is already utilized by major Google teams, including Google Search and YouTube. The library includes classic metrics like accuracy, precision, and recall, as well as specialized metrics for NLP and vision models, such as Perplexity, BLEU, IoU, and SSIM. Metrax leverages JAX's strengths, such as vmap and jit, to perform multiple "at K" operations efficiently. The library supports iterative evaluations with its merge function, facilitating the aggregation of metrics over training runs. Metrax is open to community contributions on GitHub, with some existing metrics added by contributors, and is part of the broader JAX ecosystem, which offers additional libraries and resources for building machine learning models.

Agent Garden has been made available to all users, expanding beyond just Google Cloud users, to facilitate the creation of AI agents, especially sophisticated multi-agent systems. It addresses challenges like developing AI solutions for complex business issues and integrating agents into existing systems by offering a repository of curated samples and tools. With the Agent Development Kit (ADK), developers can explore code samples that integrate with cloud services such as BigQuery and Vertex AI Search. The platform allows for easy deployment of sample agents with one click through the Agent Starter Pack and offers customization options via Firebase Studio. The introduction of Agent Garden aims to make the process of developing, deploying, and customizing AI agents seamless, as demonstrated by early adopters like Renault Group, which successfully integrated a data scientist agent into their EV charger platform.

The newly introduced User Simulation feature in the Agent Development Kit (ADK) aims to streamline the testing process for conversational agents by dynamically generating user prompts to evaluate an agent's ability to fulfill user intent, rather than relying on rigid, scripted interactions. This feature, powered by a language model, allows developers to define high-level conversation scenarios, which the simulator uses to create flexible, goal-oriented conversations that are resilient to minor changes in the agent's behavior. By configuring various parameters such as model type and turn budget, developers can tailor simulations to their specific needs, resulting in a more efficient and reliable testing process that reduces maintenance effort and enhances test coverage. The User Simulation feature is part of a broader vision to provide comprehensive simulation capabilities throughout the agent lifecycle, and the ADK team encourages users to explore and provide feedback on this innovative tool.

Go has been added to the Agent Development Kit (ADK) family, allowing developers to build AI agents using Go's concurrency and strong typing for robust and scalable applications. ADK is an open-source toolkit that enables fine-grained control over AI agents by moving the complexity of orchestration, behavior, and tool-use directly into code, ensuring robust debugging, versioning, and deployment flexibility across various environments. ADK for Go provides seamless data integration with support for over 30 databases via the MCP Toolbox and aligns with the core design principles of Python and Java ADK, offering a consistent development experience. It supports modular multi-agent systems, leveraging the Agent2Agent (A2A) protocol to enable collaborative problem-solving among agents, whether local or remote, with secure interactions. Developers are encouraged to explore the new features, join the community, and utilize resources such as source code, samples, and documentation available online.