Welcome to the May 16–31, 2026 edition of Google Cloud Platform Technology Nuggets. The nuggets are also available on YouTube.
Infographic for the post generated by NotebookLM
AI and Machine Learning
If you are looking to stay updated on the latest Google Cloud AI tools, you should bookmark this link that covers updates made across the months.
There has been a spate of announcements at Google Cloud Next 2026 in April and Google I/O held in May introduced several other releases too, that included:
- Gemini 3.5 Flash models
- Gemini Omni (any output from any input, starting with video)
- Google Antigravity in its new version
- Managed Agents on Agent Platform
Check out the post that highlights these announcements in detail.
Google Cloud has introduced the general availability of Nano Banana 2 (Gemini 3.1 Flash Image) and Nano Banana Pro (Gemini 3 Pro Image) on the Gemini Enterprise Agent Platform and the Gemini API. These enterprise-grade models allow developers to integrate image generation and editing capabilities directly into their applications and workflows.
Long-running agent workflows is where the industry is moving. But how to manage and operate them reliably and efficiently. Google believes you can use Agent Executor, an open-source runtime standard designed for executing, resuming, and deploying distributed AI agents. The key technical capabilities of this standard include:
- Durable execution to automatically resume progress via event logs and snapshotting after human-in-the-loop steps or outages
- Secure isolation through sandboxes to safely execute code and manage multi-tenant workloads
- Session consistency using a single-writer architecture to prevent state corruption
- Connection recovery to allow disconnected clients to reconnect and backfill missed responses
- Trajectory branching to let agents test alternative decision paths using checkpoints.
Check out the blog post for more details.
Looking to deploy and optimize large language models on mobile devices, take a look at new benchmarking and debugging capabilities within the Google AI Edge Portal. The platform allows developers to run automated testing across a fleet of over 120 Android device types to evaluate performance across CPU and GPU backends using the LiteRT-LM format. Check out the blog post.
Data Analytics
Data Agent Kit is an open-source collection of data engineering and data science skills, Model Context Protocol (MCP) tools, and plugins designed to connect development environments like VS Code and various CLIs with enterprise data. The kit provides pre-codified skills for tasks like query optimization, data validation, and troubleshooting, alongside MCP tools that create secure connections to platforms like BigQuery, AlloyDB, and Google Cloud Storage without requiring manual pipeline code. Check it out.
Remote Model Context Protocol (MCP) Server for AlloyDB is now in General Availability. The Remote MCP Server for AlloyDB runs on fully-managed Google Cloud infrastructure and exposes an HTTP endpoint that connects your AI applications to your data. It is integrated with the Agent Registry, supports fine-grained authorization, allows agents to not just run queries but also manage instances and more. Check out the blog post that highlights the features and a demo of how an agent has been integrate to work with AlloyDB.
Connected Sheets, which provides a direct connection between Google Sheets and BigQuery, has been there for a while. Instead of exporting data as CSVs, which causes version control and security risks, this feature allows users to perform ad-hoc analysis, modeling, and reporting on live data without writing SQL. Check out a blog post, which gives you a quick overview of Connected Sheets, walk through real-world use cases, and show you how to perform enterprise-grade data analysis using BigQuery directly in Google Sheets.
Databases
AlloyDB has got a significant upgrade to its High Availability (HA) architecture: Hot Standby. With the new architecture, the standby node continuously streams and applies write-ahead logs (WAL) from the primary node. This change allows the system to detect failures and promote the standby node within 30 seconds, dropping the average failover time to approximately 15 seconds. Furthermore, because the standby node constantly replays logs, its memory caches stay warm, allowing the database to maintain its transaction-per-second rate immediately after a failover without any performance brownouts. This feature is automatically rolled out to new instances on PostgreSQL 18 at no additional cost, with earlier versions receiving support in the following months. For more details, check out the blog post.
Developers & Practitioners
You are a developer working on building out Agents on Google technologies. How do you know which one to pick to develop these agents, since there are quite a few ranging from no-code to heavily code-driven frameworks. The article is an excellent way to step back and look at the existing landscape today to develop agents on Google Cloud, with options ranging from low code (Agent Studio) to ADK 2.0. Included in the spectrum is the recently announced Managed Agents API, which allows you to define agentic behavior and let Google Cloud handle the heavy lifting, acting as an agent-as-a-service with nothing to manage.
It is given that we need to secure our credentials, especially API Keys that may have been configured to have access to several Google Cloud services and could be a problem, if misused. But what are the recommended best practices to manage and secure your API Keys. This post goes into the detail while highlighting services like Secret Manager to manage/rotate keys, regular audits and more.
Reducing friction for developers trying out developer tools from Google and eventually deploying them on Google, has traditionally faced the common struggle of folks signing up for billing first. With a newly announced deeper integration between Google AI Studio and Google Cloud, you can now vibe code with Google AI Studio with a possible persistent layer backed by Firestore and deploy two applications on Google Cloud Run, without the need of a credit card. Check this out.
Gemini Live Agent Challenge has announced the winners and key highlights from the contest that asked developers to integrate real-time multimodal capabilities using the Gemini Live API, the Agent Development Kit (ADK), and Google Cloud infrastructure. The global challenge resulted in projects across three main categories:
- Live Agent (Voice)
- Creative Storyteller (Audio / Visual)
- UI Navigator (Gesture-driven interactions)
Notable winners included ORION, a voice-directed surgical co-pilot for robotic surgery; drone-copilot, which replaces manual joysticks with natural language voice commands for navigation and autonomous visual inspections; and Moonwalk, a hands-free desktop assistant that automates workflows by controlling the keyboard and mouse through voice.
AI agent conversations usually go back and forth with a series of questions and usually the interactions end up with just text data and interpreting the intent and extracting data can be not just challenging but more efficient in terms of number of interactions. What if a set of questions that the Agent wants to ask of the user, is converted into a simple UI with the right kind of input controls to select the data from. For e.g. a dropdown, a few checkboxes, etc. Enter A2UI (Agent to UI) Protocol, an open protocol for agent-driven user interfaces and a great example of how this protocol can be embedded into the front-door Gemini Enterprise application is demonstrated in this blog post.
Customers
One of the interesting blog posts that I look forward to each month is one in which we can learn what customers have been building with Google Cloud. For the month of May, there were interesting things that they built. Some of which include:
- Glance automated its short-form video extraction pipeline by combining Google Cloud Speech-to-Text v2, Gemini 2.5 Flash, the Google Vision API, Samurai object tracking, OpenCV, and MoviePy.
- Urban Outfitters migrated its 11TB Oracle database to AlloyDB for PostgreSQL to power its IBM Sterling OMS, deploying two read replicas to lower data latency for reporting and analytics.
- Movix engineered an agentic AI solution for dental lab quality control using the Gemini Enterprise Agent Platform alongside Cloud Run with L4 GPUs and Compute Engine VMs.
Check out the post for more of what cool things that customers built.
Containers and Kubernetes
Google Kubernetes Engine (GKE) Agent Sandbox helps you manage isolated, stateful, and single-replica workloads on GKE. It is optimized for use cases like AI agent runtimes, where untrusted, LLM-generated code must be executed in a secure and performant environment. The GKE Agent Sandbox is now generally available and there is a new introduction too, Agent Substrate. This open source project introduces a minimal control plane that moves agents on and off ready compute capacity in real-time while integrating data locality directly into its scheduler to minimize overhead. . Check out the blog post.
Security and Identity
If you are looking to protect your systems from high-speed, automated cybersecurity threats, Google Cloud has introduced Google AI Threat Defense, an autonomous platform built on a four-step framework: Prepare, Scan and Prioritize, Remediate, and Monitor. The system reduces attack surface exposure through Wiz, which maps your live environment and uses an AI penetration testing agent to validate exploitable pathways. When vulnerabilities are discovered, the platform incorporates Mandiant expertise for response planning while CodeMender automatically generates fixes, analyzes library dependencies, and creates verification tests directly inside the developer’s integrated development environment or command-line interface. Check out the blog post for more details.
The second Cloud CISO Perspectives for May 2026 is out. If you are looking to build an AI-ready security program for the public sector, this article outlines a structured, 12-month roadmap designed to manage complex systems and reduce operational toil. The plan is organized across five core workload domains: executive alignment, process optimization, talent augmentation, posture elevation, and advanced governance.
Infrastructure and Networking
There have been several architectural updates to Google’s global and data center networks designed to handle the specific traffic patterns and scale required for AI training and inference workloads. The new Virgo Network serves as a flat, two-layer scale-out data center fabric that connects up to 134,000 eighth-generation TPU chips, delivering increased bandwidth, lower latency, and autonomous reliability features like automated hang and straggler detection to isolate faulty instances. Read more here.
Given the kind of infrastructure that Google runs, it should be interesting to learn about how they ensure that any optimizations that they introduce to their infrastrcture takes place safely. In fact they run a fleet-wide, machine-level A/B experimentation framework. Instead of testing at the application level, Google enables changes on individual machines across a balanced 1% subset of the fleet to capture system-wide effects for core components like libraries, compilers, kernels, and cluster management systems. Check out the details here.
DevOps and SRE
SRE AI is an initiative that integrates agentic AI into Site Reliability Engineering. This includes deploying autonomous AI agents across the entire software development lifecycle. These agents assist in reliability design by creating and updating playbooks, use models like TimesFM for anomaly detection and alert enrichment, and orchestrate incident management by summarizing communications and drafting postmortems. For more details, check out the blog post.
Application Development
Google Cloud has introduced App-centric maintenance visibility within Unified Maintenance to shift planned maintenance tracking from an infrastructure-focused view to a business-oriented view. By integrating directly with App Hub, Unified Maintenance automatically aggregates the maintenance schedules of registered resources (GKE clusters, GCE VMs, or AlloyDB instances) into a single dashboard centered around the application as the primary unit of management. To use this feature, teams need to enable the Maintenance API and configure application boundaries in the Google Cloud Console. For more details, check out the blog post.
Continuing on Application Development, we now have a public preview of AppLifecycle Manager Feature Flags (ALM FF). This service decouples feature releases from code deployments, allowing teams to ship code with new features disabled by default and use a toggle as an instant kill switch if issues arise. Built on the open-source OpenFeature standard and the flagd evaluation engine, ALM FF uses Common Expression Language (CEL) to enable targeted rollouts, such as percentage-based traffic ramping or precise allowlisting for specific internal teams and testers. Check out the blog post.
Learn about Google Cloud
If you have been working with traditional transactional databases and are looking to transition enterprise data from static reports to autonomous systems, Google Cloud outlines a structural evolution across five technical scenarios to manage security, costs, and accuracy. These five scenarios are:
- Scenario 1: The Static API Contract, which uses pre-written, parameterized SQL queries for deterministic performance.
- Scenario 2: Custom Agent with SQL Generation, where an LLM translates natural language into queries using schema metadata.
- Scenario 3: Conversational Analytics, which uses a platform-native engine grounded in verified query libraries to enforce business logic.
- Scenario 4: Managed MCP Tools introduces the open-source Model Context Protocol (MCP) to decouple the reasoning layer from tool execution via a managed BigQuery server. This helps to scale across different systems.
- Scenario 5: Custom Hosted MCP Servers gives engineers full control to build specialized, multi-source tools on infrastructure like Cloud Run, allowing for programmatic governance and custom data abstraction.
Check out the blog post that provides detailed guidance, queries and more.
Write for Google Cloud Medium publication
If you would like to share your Google Cloud expertise with your fellow practitioners, consider becoming an author for Google Cloud Medium publication. Reach out to me via comments and/or fill out thisform and I’ll be happy to add you as a writer.
Stay in Touch
Have questions, comments, or other feedback on this newsletter? Please send Feedback.
If any of your peers are interested in receiving this newsletter, send them theSubscribe link.
Top comments (0)