March 2026 Summaries

The Agent Development Kit for Go 1.0 represents a significant advancement in the deployment of AI agents from experimental scripts to robust production services, emphasizing observability, security, and extensibility. With the introduction of features like OpenTelemetry integration for tracing, a Plugin System for self-healing logic, and Human-in-the-Loop confirmations for sensitive operations, the update aims to address challenges such as non-determinism and agent failures. The integration of YAML configurations allows for easier agent management and ensures cross-language consistency, facilitating rapid iteration on agent personas without altering core binaries. Additionally, the refined Agent2Agent protocol ensures stable communication between agents across different programming languages like Go, Java, and Python, thereby enhancing task delegation and insight sharing. This release invites developers to explore and contribute to this evolving landscape by engaging with resources such as a Quickstart Guide and community platforms.

The continuous checkpointing feature in Orbax and MaxText aims to optimize the balance between reliability and performance during model training by generating checkpoints more dynamically rather than at fixed intervals. Traditional fixed checkpointing can lead to inefficiencies, either through infrequent checkpoints that risk data loss during failures or too frequent checkpoints that can bottleneck performance. Continuous checkpointing minimizes these risks by asynchronously saving checkpoints only after the previous save operation completes, thereby maximizing resource utilization without significant performance degradation. This approach is particularly beneficial for large-scale training tasks, as it reduces device-to-host blocking times and scales efficiently with the size of operations. The system also allows for customizable policies to further tailor checkpointing behavior to specific use cases, promoting efficient resource conservation and management. However, the effectiveness of continuous checkpointing is highly dependent on network bandwidth, emphasizing the importance of co-locating storage with the training cluster to avoid cross-metro network delays that could compromise reliability.

Google has released version 1.0.0 of its open-source Agent Development Kit (ADK) for Java, expanding its multi-language ecosystem that began with Python and now includes Java, Go, and TypeScript. The latest version introduces several enhancements such as new tools like GoogleMapsTool and UrlContextTool for enhanced data grounding, a centralized plugin architecture for global execution control, and improved context engineering with event compaction capabilities. Additionally, the release supports Human-in-the-Loop workflows for agent intervention, robust session and memory management services, and native support for the Agent2Agent protocol to enable seamless collaboration between remote agents across different frameworks. These features aim to create more powerful, interoperable agents that can interact with external environments and maintain long-term conversational memory, while providing developers with greater control over agent behavior and execution.

Large language models (LLMs) have limitations in keeping up with the rapidly evolving field of software engineering due to their fixed training knowledge, leading to a knowledge gap that Google DeepMind seeks to address with agent skills. These skills offer a lightweight solution to guide coding agents using the Gemini API by providing high-level API features, current models, SDKs, sample code, and documentation entry points, encouraging real-time information retrieval. An evaluation was conducted using a skill developed for the Gemini API, revealing significant performance improvements in code generation tasks for the latest Gemini 3 series models, although older models also benefited to a lesser extent. Despite the successes, challenges remain such as maintaining up-to-date skill information and exploring better ways to incorporate live knowledge, prompting ongoing adjustments and explorations for skill enhancement.

Vibe-coding with Gemini in conjunction with MediaPipe offers developers the tools to create highly interactive games and apps that leverage real-time input control through machine learning solutions for vision, audio, and language. By utilizing Google AI Studio, developers can swiftly transform ideas into playable experiences by integrating MediaPipe's capabilities such as face, hand, and pose tracking, enabling the creation of apps that interact seamlessly with the physical world. Examples include a motion-controlled Chrome Dino game and a hair recoloring app, both demonstrating the potential of MediaPipe's real-time on-device processing to deliver responsive and immersive experiences. AI Studio facilitates this creative process by allowing iterative refinements through natural language prompts, enabling developers to enhance their applications continuously. The combination of Gemini's intelligence and MediaPipe’s suite of ML solutions allows for the development of sophisticated applications that react in real-time, paving the way for innovative digital experiences.

Extracting text from unstructured documents has traditionally been challenging, especially for complex layouts like multi-column PDFs and nested tables, but advancements in large language models (LLMs) are now facilitating reliable document understanding. LlamaParse is a tool that enhances traditional Optical Character Recognition (OCR) by integrating multimodal capabilities and customized parsing instructions, improving text extraction from documents such as PDFs, presentations, and images. Utilizing Gemini 3.1 Pro, LlamaParse offers a robust workflow for parsing brokerage statements, which include dense financial jargon and complex tables, by breaking down the process into stages: ingesting the document, routing and parsing it, extracting text and tables concurrently, and synthesizing a summary using Gemini's capabilities. This approach not only ensures high-quality parsing but also optimizes for both accuracy and cost by using a two-model architecture. The setup involves installing necessary Python packages and configuring API keys, followed by creating a LlamaParse client to parse and extract information from documents. The workflow is designed to be scalable and resilient, running text and table extractions in parallel to minimize latency and enhance performance. This system demonstrates how combining LLM capabilities with dedicated parsing tools can structure complex data, making applications like personal finance assistants more efficient and reliable.

The text explores the complex landscape of AI agent development, emphasizing the challenges and solutions associated with various protocols like MCP, A2A, UCP, AP2, A2UI, and AG-UI. These protocols aim to streamline the integration of AI agents with tools, APIs, and frontend components, saving developers from writing extensive custom code. By using the Agent Development Kit (ADK), the text illustrates how a multi-step supply chain agent for a restaurant can be built, starting with a basic LLM and progressively incorporating protocols to manage inventory, place orders, authorize payments, and render dashboards. Each protocol addresses specific needs: MCP for connecting agents to data, A2A for agent-to-agent communication, UCP for standardizing commerce transactions, AP2 for securing payment authorizations, A2UI for creating flexible interfaces, and AG-UI for streaming interactions. The text encourages adopting these protocols to build a robust, scalable, and interactive AI agent ecosystem while leveraging existing ADK tools and integrations.

The integration of AI agents with Google Colab through the new, open-source Colab MCP (Model Context Protocol) Server allows for enhanced prototyping by leveraging Colab's cloud environment, freeing users from the limitations of local machines. This setup enables AI agents to programmatically control the Colab notebook interface, facilitating the automation of tasks such as creating data analyses, managing dependencies, and organizing content within notebooks. Users can treat Colab as an automated workspace, with agents capable of executing Python code, generating visualizations, and producing fully reproducible, executable artifacts in the cloud. The Colab MCP Server is designed to remove the friction between local development environments and cloud computing, encouraging user feedback and community involvement to refine and expand its capabilities.

Plan mode is a new feature introduced to the Gemini CLI, designed to analyze requests and plan complex changes in a safe, read-only environment, preventing accidental code modifications. It allows users to explore codebases, search patterns, and refine strategies by using specialized tools like codebase_investigator and the ask_user tool for bi-directional communication, ensuring plans align with user goals. Plan mode supports integration with external data sources through MCP tools and can be extended with custom policies. The Conductor extension enhances plan mode by facilitating complex workflows and context-driven development, ensuring a thorough pre-flight check for multi-step projects. Users can customize their experience by enabling or disabling plan mode based on their preference, with options for starting sessions in research-first mode or opting for Auto-Edit mode for direct implementation. Feedback is encouraged to enhance the feature, with updates available through Gemini CLI's GitHub and social media channels.

Gemini Code Assist extensions for IntelliJ and Visual Studio Code have introduced two new features, Finish Changes and Outlines, to improve the developer workflow by reducing friction and enhancing code comprehension. Finish Changes acts as an AI pair programmer, completing code modifications based on partial inputs like pseudocode or comments, and applying repetitive changes across files while integrating project-specific styles. This feature allows developers to maintain their flow state without the need for extensive prompt engineering. Meanwhile, Outlines offer concise, high-level summaries interleaved with the source code, providing a living design document that helps developers understand complex codebases quickly. This feature includes interactive navigation and customizable views, making it easier to keep documentation synchronized with code changes. Together, these tools aim to enhance developer efficiency by simplifying code comprehension and implementation processes.

Gemini Code Assist has introduced several enhancements and new features to improve the coding experience by making it more fluid and efficient for developers. These updates include Agent Mode with Auto Approve, which acts as an intelligent collaborator simplifying large-scale code changes, and Inline Diff Views for a dynamic code review process with AI-generated suggestions. The Revert to Checkpoint feature allows developers to experiment fearlessly by easily rolling back changes, while multi-part chat code suggestions provide granular control over AI-generated modifications. Improvements in code completion speed and Chat Code Suggestion Preview enhance workflow efficiency by maintaining cognitive flow and providing clear, manageable code suggestions. The Context Drawer and customizable commands enable developers to tailor the tool to their specific needs, ensuring relevant assistance and automating routine tasks. Additionally, features such as adding code snippets, terminal output to chat context, and configurable codebase awareness enhance the precision of AI-assisted coding by focusing on relevant parts of the codebase. These changes collectively aim to transform Gemini Code Assist into a highly specialized tool that adapts to individual developer workflows, fostering a collaborative and intuitive software development environment.

Wednesday Build Hour is a weekly event designed for developers, cloud architects, and engineers to engage in a live, interactive learning experience led by Google Cloud experts. It serves as a dedicated time for participants to enhance their technical skills by working on practical projects rather than passively consuming information. The sessions cover a wide range of topics, including legacy modernization, AI advancements with tools like Vertex AI, and development productivity enhancements using Google Antigravity. Participants have the opportunity to explore cutting-edge technologies, such as AI agents and multimodal AI, while connecting with a community of like-minded professionals. The recurring nature of the event encourages the formation of a consistent learning habit, allowing attendees to integrate new knowledge into their workflows effectively.

TensorFlow 2.21 introduces significant updates, most notably the formal release of the LiteRT stack, which boasts 1.4x faster GPU performance and new NPU acceleration, enhancing its status as a universal on-device inference framework. This advancement simplifies the workflow for GPU and NPU acceleration across edge platforms, supports cross-platform GenAI deployment, and offers seamless model conversion with PyTorch and JAX, all while maintaining reliable cross-platform deployment. The update also includes extended support for lower-precision data types to improve performance and efficiency across various operators. Additionally, TensorFlow commits to addressing community feedback by increasing efforts towards security, bug fixes, and timely dependency updates, while continuing to engage with community contributions. The TF Lite project has been rebranded as LiteRT, and TensorFlow recommends checking out the latest updates for Keras 3, JAX, and PyTorch for developments in Generative AI.

Google Cloud Next '26, taking place in Las Vegas from April 22-24, is the largest annual event for developers working on Google Cloud, focusing on the latest advancements in agentic AI, modern infrastructure, and collaborative problem-solving. Attendees will have the opportunity to engage in live coding sessions and technical breakouts, exploring topics such as AI-driven developer workflows, real-time multimodal applications, and zero-trust security architecture. The event emphasizes the unique value of in-person interactions, offering numerous meetups and networking opportunities to foster collaboration and idea exchange among developers and industry experts. Highlighted sessions include insights into scaling agentic applications, optimizing infrastructure for global enterprises, and enhancing observability and deployment efficiency. While many sessions will be available online, the organizers stress that the spontaneous connections and innovative discussions are best experienced in person.

Google I/O 2026, scheduled for May 19-20 at the Shoreline Amphitheatre in Mountain View, CA, is set to highlight the potential of AI in enhancing developer workflows through an engaging pre-event puzzle. This year's puzzle comprises five games across various genres, with a hidden sixth game, and showcases the collaborative efforts between creatives and developers using Google AI Studio and Google Antigravity. The Gemini models play a pivotal role in the creative and developmental processes, enabling dynamic level generation and integrating AI-driven gameplay features such as an AI caddy and voice-controlled Android Bot. Upon completing the games, players unlock a virtual pet, Dino Pal, with adaptive personality traits. A unique aspect of the puzzle is the invitation for participants to explore and remix the underlying code in Google AI Studio, encouraging creativity and innovation, with completed projects shared via #GoogleIO. Registration for updates is available at io.google.