Skip to main content
AI Learning Resources

Google Cloud Tech

@googlecloudtech 1M subscribers73 videos

Official Google Cloud channel - Vertex AI, Gemini integrations, and AI/ML engineering tutorials.

  • official
  • google
  • cloud
  • vertexai
Visit channel

Latest videos

Thumbnail for What happens to your AI agent when you close your lid?0:56

What happens to your AI agent when you close your lid?

Traditional agentic workflows depend on a specific open IDE window, browser tab, or active local terminal session. Closing a laptop lid, putting a machine to sleep, or experiencing a network interruption causes locally hosted AI agent tasks to crash. Persistent cloud actors decouple the execution runtime from the client device so the agent stays active regardless of user connectivity. Cloud-hosted AI environments can be accessed and controlled through diverse interfaces including Slack, command-line tools, mobile phones, and Google Docs.

Oct 23, 2026Read summary
Thumbnail for Do more with fewer tokens: Gemini Smart Routing0:33

Do more with fewer tokens: Gemini Smart Routing

Development teams building artificial intelligence agents frequently face operational roadblocks due to token limits. Google Cloud introduced Smart Routing with Gemini to help manage token usage and optimize model efficiency. The system allows users to define a hard, real-time spending cap at a granular level. Sub-tasks are automatically routed to the fastest and most efficient model capable of completing the work. Organizations must balance token consumption with agentic success when deploying AI systems.

Oct 23, 2026Read summary
Thumbnail for How smart storage turns dark data into AI context1:00

How smart storage turns dark data into AI context

Most enterprise information exists as dark data consisting of unstructured files like PDFs, scans, and media scattered across servers. Organizing scattered business records and meeting notes often requires extensive manual effort or dedicated full-time work. Making dark data usable for AI systems traditionally requires engineering teams to construct fragile extract, transform, and load pipelines. Smart Storage identifies and classifies unstructured data directly in place where it already lives. Pairing Smart Storage with Knowledge Catalog maps disorganized data so artificial intelligence agents can reason over it.

Oct 22, 2026Read summary
Thumbnail for Plan your app's system spec (No code)24:08

Plan your app's system spec (No code)

Transitioning from a prototype to a production-ready application requires deliberately specifying data models, server contracts, and agent behaviors before writing code. Using a pre-populated static exercise library alongside generative AI reduces latency from fifteen seconds to two seconds while cutting token costs. Software architecture should follow the principle of never storing derived data that can easily be calculated by client code. A staged mutation pattern protects databases by requiring user approval through inline cards before any AI-generated change commits to storage. A tiered memory model categorizes user data into permanent staples, time-expiring physiological notes, and situational context. Weekly chat buckets balance Firestore reading efficiency against the one-megabyte document limit when storing conversation logs.

Oct 22, 2026Read summary
Thumbnail for Which model should I use? | The Agent Factory23:10

Which model should I use? | The Agent Factory

Developers can select models automatically using agent routing, quantitatively using formal evaluations, or intuitively by running comparative test prompts across models. Using declarative, text-based outputs like Strudel live-coding music markup or Mermaid diagram syntax allows humans to manually edit and refine model outputs without full regeneration. Smaller models such as Gemini Flash Lite are often sufficient, faster, and more cost-effective for straightforward tasks like generating short ambient music tracks or basic sequence diagrams. Advanced and multi-step tasks, including video composition in HyperFrames and complex polyphonic music transcription, benefit significantly from the reasoning capabilities of Gemini 4 Argon. Model outputs across varying sizes can be combined in production, such as taking lower-third styling generated by Flash and pairing it with a layout created by Gemini 4 Argon.

Oct 21, 2026Read summary
Thumbnail for AI coding agent on production database? True or False?0:57

AI coding agent on production database? True or False?

Granting an AI coding agent direct access to a production database exposes software systems to severe operational and legal hazards. A single hallucination from an AI coding agent can generate unintended deletion commands that erase an entire database. Feeding live personal records to an AI model risks violating data privacy regulations and incurring regulatory fines. Developers must sandbox AI coding agents so they are isolated from production environments and restricted to dedicated development databases. Any production database copy supplied to a development environment must be sanitized by substituting real personal records with fake data.

Oct 19, 2026Read summary
Thumbnail for How to build and deploy powerful data apps with agents1:38

How to build and deploy powerful data apps with agents

Data applications can be hosted directly inside Looker while preserving the governance and security of the LookML semantic layer. Initial setup requires executing a terminal installation script and providing a Looker instance URL, API credentials, and a target folder ID. Developers can use AI coding agents such as Antigravity or Claude Code to generate multi-tab user interfaces from natural language prompts. Applications can be previewed locally, customized using uploaded images or logos for color themes, and published to Looker with a single prompt. Deploying data apps enables JavaScript interactivity and custom visualizations while querying underlying Looker explores.

Oct 16, 2026Read summary
Thumbnail for IDE-driven deployments: Creating & debugging pipelines with Harness and Google Cloud2:50

IDE-driven deployments: Creating & debugging pipelines with Harness and Google Cloud

Harness integrates with Google Cloud Antigravity IDE and Gemini models to enable pipeline creation and troubleshooting directly inside developer environments. Developers can generate comprehensive CI/CD pipelines, including Docker builds, security scanning, approval gates, and Kubernetes canary deployments, using natural language prompts. Pipeline configurations authored through the IDE are translated into declarative YAML and visualized within the Harness web console. Gemini can analyze past deployment failures to identify root causes, such as missing pipeline parameters, and suggest fixes without leaving the editor. The integration provides high-level diagnostics across multiple recent deployments to surface common failure trends across microservices and cluster environments.

Oct 9, 2026Read summary
Thumbnail for Run, hunt, hack: Join the Cloud Run hackathon on product hunt0:54

Run, hunt, hack: Join the Cloud Run hackathon on product hunt

Google Cloud is hosting a Cloud Run Hackathon on Product Hunt starting on Wednesday, October 14th. Developers can deploy an app, website, or game to Cloud Run using a single command. Cloud Run manages infrastructure automatically by scaling based on incoming traffic volume. Participants can receive 50 dollars in Google Cloud credit upon signing up, while supplies last. Submitted projects deployed to Cloud Run and launched on Product Hunt are eligible to win prizes including 500 dollars in Google Cloud credits. Product Hunt forums offer participants access to chat with Google engineers and Google Developer Experts.

Oct 9, 2026Read summary
Thumbnail for Navigating AI automation in Looker: Essential developer tools for agents5:55

Navigating AI automation in Looker: Essential developer tools for agents

Looker Skills are Markdown instruction files that allow AI agents to dynamically retrieve domain-specific rules on demand while keeping context windows clean. The open-source, Go-based Looker Command Line Interface wraps the Looker API to allow AI agents to manage platform resources like dashboards, folders, and schedules. Administrators should pair the Looker Command Line Interface with dedicated service accounts bound to least privilege rather than personal user credentials. The Looker Managed Model Context Protocol Server provides native, instance-wide tool execution directly inside Looker without requiring external hosting infrastructure. The open-source Model Context Protocol Toolbox for Databases allows self-hosted multi-source querying across Looker and database engines such as BigQuery and AlloyDB. The Looker Visual Studio Code extension synchronizes local development mode with Looker while providing a local proxy to connect AI agents to the managed Model Context Protocol server.

Oct 8, 2026Read summary
Thumbnail for Build & connect an incident investigation agent with Elastic Agent Builder and Gemini Enterprise3:56

Build & connect an incident investigation agent with Elastic Agent Builder and Gemini Enterprise

Elastic Agent Builder enables developers to create specialized AI agents equipped with tools and skills such as investigation runbooks. Observability data including logs, metrics, and traces can ground an AI agent to automatically detect and isolate system regressions. Agents created in Elastic Agent Builder can be configured via a graphical interface or programmatically through an API. The Agent-to-Agent protocol allows custom Elastic agents to be registered and invoked directly within Gemini Enterprise. Authentication currently uses an intermediate proxy on Cloud Run to manage API keys, with plans to support OAuth in the future.

Oct 7, 2026Read summary
Thumbnail for Chrissie Goodrich takes over the Google Cloud Tech channel!0:29

Chrissie Goodrich takes over the Google Cloud Tech channel!

Chrissie Goodrich is hosting a month-long takeover on the Google Cloud Tech channel. The takeover content focuses comprehensively on data architecture, tooling, and analytics. Specific topics scheduled for discussion include data visualizations, data agents, and data pipelines. The series will also cover databases, semantic layers, and business integrations. Viewers are prompted to subscribe to the channel to follow upcoming takeover updates.

Oct 7, 2026Read summary
Thumbnail for BigQuery Graph measures explained (Why your AI agents get metrics wrong and how measures fix it)8:48

BigQuery Graph measures explained (Why your AI agents get metrics wrong and how measures fix it)

Joining relational tables across many-to-many relationships duplicates rows and causes standard SQL aggregate functions to inflate numerical metrics. BigQuery Graph measures allow calculations to be defined at the data model level rather than rewritten inside individual SQL queries. A graph measure uses entity keys to ensure values like stream counts are calculated once per entity within an aggregation group. The AGG function applies graph-defined measure rules within queries using GRAPH_EXPAND, regardless of how groupings or dimensions change. Graph structures preserve relationships between entities, enabling topological impact analyses such as determining exposure to specific playlist nodes. Defined property graphs and measures integrate into Knowledge Catalog and can be queried directly by BigQuery conversational analytics agents.

Oct 6, 2026Read summary
Thumbnail for Agent Evaluation walkthrough0:25

Agent Evaluation walkthrough

The evaluation interface includes a visual score heatmap that maps out agent test performance across individual evaluation questions. Deterministic metrics and specific question-by-question deep dives allow developers to inspect where test runs passed or failed. Radar charts display multi-dimensional LLM judge metrics, including groundedness, citation fidelity, and markdown formatting structure. The dashboard tracks evaluation iterations over time and generates automated technical diagnostic reports based on the run results.

Oct 5, 2026Read summary
Thumbnail for How Antigravity Evaluates a LangGraph Agent0:29

How Antigravity Evaluates a LangGraph Agent

The AI assistant automatically generates an architecture overview diagram outlining the evaluation workflow. The implementation plan configures OpenInference trace conversion for LangGraph agents into canonical evaluation datasets. The tool automatically detects local installations of agent-eval and the DocsHound repository. A local DocsHound agent runs background traces to produce interaction data for evaluation metrics.

Oct 5, 2026Read summary
Thumbnail for How to build an agentic C-Suite with Google Cloud ADK and Cloud Run6:26

How to build an agentic C-Suite with Google Cloud ADK and Cloud Run

Google Developer Expert Anupam Phoghat built an automated C-suite multi-agent system powered by Gemini models and the open-source Agent Development Kit (ADK). The application assigns executive roles such as CSO, CTO, CMO, and CFO to specialized agents coordinated by a central orchestrator. A human-in-the-loop review mechanism allows operators to verify, tweak, or reject intermediate agent findings before final recommendations are compiled. The multi-agent workflow runs serverless on Google Cloud Run, incurring zero baseline costs when idle. Running five to six monthly strategic workflows producing roughly twenty outputs costs approximately seven to ten dollars per month.

Oct 1, 2026Read summary
Thumbnail for How to build AI agent evals in 4 steps0:55

How to build AI agent evals in 4 steps

Evaluating an AI agent begins by connecting Antigravity to the agent source code to understand its internal architecture. A standardized evaluation toolkit can reduce evaluation setup time from several weeks to a single hour. Developers can define expected quality outcomes in configuration files to generate targeted evaluation metrics. Running automated evaluations produces quantified metrics and visual reports that guide iterative agent improvements.

Oct 1, 2026Read summary
Thumbnail for How to use AI Agents for automated code reviews0:44

How to use AI Agents for automated code reviews

Harness engineering consists of sophisticated techniques designed to keep an AI agent aligned with its baseline. Small pull requests simplify the review process, making it feasible to deploy agents for automated reviewing. Restricting the state space of small pull requests enables agents to operate over longer horizons with less human intervention. Scaling agent autonomy relies on building confidence in small changes and stacking them end to end over a longer arc. Increasing the operational loop allows human supervisors to task agent systems with more ambitious development goals.

Sep 30, 2026Read summary
Thumbnail for Why Your AI Agent Fails in Production (And How to Catch It)26:08

Why Your AI Agent Fails in Production (And How to Catch It)

Evaluating AI agents with measurable metrics prevents silent failures and performance regressions before deployment to production environments. OpenTelemetry and OpenInference standards enable uniform capture of agent traces, spans, and metadata across different agentic frameworks like LangGraph and Google ADK. Antigravity can inspect agent source code and runtime traces to generate architecture diagrams and reproducible evaluation datasets from existing GitHub repositories. Evaluation suites require combining deterministic metrics such as latency and token costs with LLM-as-a-judge criteria like markdown formatting, citation fidelity, and grounding. Visual evaluation reports provide actionable score heatmaps that reveal hidden flaws in documentation generation and guide targeted prompt improvements.

Sep 30, 2026Read summary
Thumbnail for Harness engineering and context curation1:26

Harness engineering and context curation

Ryan Lopopolo aims to be an intentionally lazy prompter by shifting effort from writing detailed prompts to curating the environment. Lopopolo notes he has avoided manually opening a code editor since approximately May of the previous year. AI models improve over time, but they do not automatically know user preferences without explicit contextual grounding. Harness engineering focuses on curating context and environment around the AI model so it understands developer requirements upon booting up. Providing tools and environmental context directly to the model reduces the need to front-load complex instructions into prompts.

Sep 29, 2026Read summary
Thumbnail for Make your AI interview YOU first1:07

Make your AI interview YOU first

The /grill-me command creates an interactive pre-code interrogation session before an AI writes software. Instead of generating code immediately from prompts, the tool asks architectural questions covering latency thresholds, cable vibration stress, and bandwidth. Adding the --strict flag triggers rapid-fire questioning on edge cases like packet drops and fuel starvation in specific corners. Engineers complete the interrogation checklist inside the terminal to nail telemetry specifications before deployment.

Sep 29, 2026Read summary
Thumbnail for What exactly is an Agent Harness?1:22

What exactly is an Agent Harness?

An AI agent consists of a large language model combined with an agent harness. The harness represents everything within an AI agent architecture that is not the language model itself. A standalone language model can answer static knowledge queries directly but requires external information for contextual or dynamic questions. When faced with a question requiring real-time data, the model instructs the harness to retrieve the necessary facts. The agent harness collects external context, combines it with the user's original query, and passes the complete prompt back to the language model to produce a final response.

Sep 28, 2026Read summary
Thumbnail for A Developer's Guide to Gemini 3.5 Transcribe5:08

A Developer's Guide to Gemini 3.5 Transcribe

Gemini 3.5 Transcribe is a specialized speech-to-text model designed for audio inputs such as customer calls, meetings, voice notes, and dictation. Developers can integrate transcription via the Google Gen AI SDK using two distinct APIs: synchronous generateContent and the real-time streaming Live API. Synchronous transcription supports audio inputs via direct byte arrays or Google Cloud Storage URIs. Configuration options in the AudioTranscriptionConfig include word-level timestamps, pinned or auto-detected language codes, speaker diarization, and custom vocabulary lists. Real-time streaming over WebSockets uses concurrent asynchronous tasks to stream raw PCM audio chunks while receiving interim and finalized transcription message types.

Sep 24, 2026Read summary
Thumbnail for Turn Audio into Action with Gemini 3.5 Transcribe5:27

Turn Audio into Action with Gemini 3.5 Transcribe

Gemini 3.5 Transcribe is Google's transcription model available through Agent Platform on Google Cloud. The model comes in two variants: Gemini 3.5 Transcribe for existing recorded audio files and Gemini 3.5 Transcribe Live for real-time streaming audio. Built-in capabilities include automatic language detection, speaker diarization, word-level timestamps, and custom vocabulary support. Custom vocabulary lists bias the model toward accurately transcribing vertical-specific terms such as medical conditions and procedures. Transcripts generated by Gemini 3.5 Transcribe can be passed directly to reasoning models like Gemini 3.7 Flash to extract sentiment scores, frustration flags, and conversation summaries.

Sep 24, 2026Read summary
Thumbnail for Prevent surprise cloud bills: enforce hard spending caps on Gemini API & Vertex AI4:55

Prevent surprise cloud bills: enforce hard spending caps on Gemini API & Vertex AI

Google Cloud allows users to set hard spending caps that automatically cut off application access before going over budget. Hard spending caps can be configured under Billing and Budgets within the Google Cloud console for services like Gemini API, Vertex AI, Cloud Run, and Cloud Run Functions. Spend caps apply to a single Google Cloud project at a time, making project isolation between development and production environments essential. When a spend cap is breached, the application receives an HTTP 403 error status code and the console displays an option to lift the cap. Spend caps trigger within minutes, but because billing continues during that window, setting budgets slightly below the absolute maximum limit is recommended.

Sep 24, 2026Read summary
Thumbnail for Gemini is coming to your city0:28

Gemini is coming to your city

Google is hosting a global event series titled Build with Gemini in cities worldwide. The events are designed for a broad audience including business leaders, architects, and app developers. Attendees can choose from three distinct tracks tailored to their specific roles and interests. The format prioritises hands-on work, moving attendees directly from the keynote into building real-world applications.

Sep 24, 2026Read summary
Thumbnail for Modernize enterprise apps on GKE: Getting started with Online Boutique2:09

Modernize enterprise apps on GKE: Getting started with Online Boutique

Online Boutique is an open-source retail store sample application designed to demonstrate how enterprise applications can be modernized on Google Kubernetes Engine. The application architecture consists of eleven microservices written across five programming languages: Python, Java, C#, Go, and Node.js. The architecture includes a dedicated front end, a load generator for simulating user traffic, and a Redis database backend. Deploying Online Boutique requires creating a GKE Autopilot cluster and applying the provided Kubernetes manifests via Google Cloud Shell. The sample repository provides quickstart instructions, extensive documentation, and optional components configured through Kustomize.

Sep 23, 2026Read summary
Thumbnail for Tools in Voice AI0:49

Tools in Voice AI

A tool execution delay of two seconds can break a voice agent conversation by making it seem as if the call dropped. Latency that goes unnoticed in text interfaces becomes disruptive during live spoken interactions. Fast tools such as play, pause, or skip should be executed immediately by the agent. Slow tools should not force the voice interface to wait, but should instead be acknowledged immediately while executing in the background. Differentiating tool execution strategies prevents voice agents from falling silent during conversations.

Sep 23, 2026Read summary
Thumbnail for Give your voice AI agent real abilities (tool calling)9:42

Give your voice AI agent real abilities (tool calling)

Language models cannot interact with external systems directly and rely on tool calling to trigger code executions during conversation. The tool calling loop consists of four distinct phases: declare, decide, execute, and return. Developers can implement tool calling either manually with the raw Gemini SDK or automatically using Python functions and docstrings in the Agent Development Kit (ADK). Live voice tools must execute synchronously and return immediately to prevent unnatural conversational silences. The before_tool_callback checkpoint allows applications to observe, block, or rewrite tool calls before execution without breaking character.

Sep 23, 2026Read summary
Thumbnail for Join Developers Building with Gemini0:23

Join Developers Building with Gemini

Google Cloud is hosting an event for developers to build generative AI systems using Gemini. Developers begin their builds by providing an initial prompt within Antigravity. The development workflow allows integration of advanced capabilities such as Memory Bank and RAG engines. Participants build alongside peers and receive guidance directly from Google experts to understand system architecture.

Sep 23, 2026Read summary
Thumbnail for Vibe coding in the pit lane1:03

Vibe coding in the pit lane

Google Antigravity allows users to attach full PDF documentation directly into prompts without manual text scraping or separate OCR scripts. A team of ten engineers built an AI racecar coach application in two weeks using Antigravity and multimodal Gemini models. The vision engine parses complex suspension vector schematics, chassis manuals, and raw sensor pinouts directly into live context. Engineers generated real-time latency fixes and telemetry adjustments between laps in the pit lane, achieving a 1:57.25 hot lap at Sonoma Raceway.

Sep 22, 2026Read summary
Thumbnail for Google Antigravity: Building a Real-Time AI Race Coach5:56

Google Antigravity: Building a Real-Time AI Race Coach

Google Developer Experts collaborated at Sonoma Raceway to build an on-device AI race coach capable of delivering low-latency real-time guidance. Because cloud connectivity was unavailable on the racetrack, the team deployed an open-weight Gemma model running directly on an onboard device Tensor Processing Unit. Vehicle telemetry including GPS, accelerometer, pedal position, and steering angle was routed through a CAN to USB converter into the AI model. The engineering team used Google Antigravity to rapidly build, debug, and iterate on the application directly trackside between 20-minute driving sessions. After initial issues with excessive chatter and mismatched track positioning were resolved, the AI coach successfully provided timely throttle prompts that increased the driver lap speed.

Sep 10, 2026Read summary
Thumbnail for Graph Engineering with ADK6:10

Graph Engineering with ADK

Relying on a single large prompt to fetch data and reason in one model call causes hallucinations because ungrounded models lack deterministic API and course data. Graph engineering organizes complex systems by separating predictable operations into deterministic Python functions while reserving large language model calls for reasoning tasks. In Google's Agent Development Kit 2.0 (ADK 2.0), functions and agent nodes operate as peers within the same orchestration graph and wiring structure. Key structural patterns in graph engineering include fan-out for parallel execution, join nodes for aggregating parallel branch outputs without custom aggregator agents, and routing for selecting execution paths. Deterministic routers using code logic are faster, cheaper, and more reliable than LLM routers whenever branching decisions are based on structured data and closed sets of options. Developers should use static graph workflows when a process structure is known beforehand, dynamic workflows when graph shape depends on runtime inputs, and autonomous agents when workflows cannot be predefined.

Sep 10, 2026Read summary
Thumbnail for Graph Engineering 1015:56

Graph Engineering 101

Graph engineering is an architecture pattern for complex artificial intelligence systems that structures processes into networks of agent nodes and deterministic function nodes. The core components of agent architectures include the harness for tools, memory, and guardrails, the loop for iterative reasoning cycles, and the graph for multi-step workflow routing. Google Agent Development Kit enables shared state across different nodes and agents throughout an orchestrated graph workflow. A pull request review pipeline illustrates graph engineering through parallel fan-out tasks, a join node that synthesizes findings, and a router that directs code to fixer agents or human approval. Graph engineering focuses on operational behavior and execution control flow, whereas knowledge graphs and GraphRAG focus on data representation. While agent swarms rely on dynamic agent personas to tackle ambiguous tasks, graph engineering offers predictability, debuggability, and control for well-defined workflows.

Sep 3, 2026Read summary
Thumbnail for How to monitor & optimize Google Gemini agents with Grafana Cloud7:22

How to monitor & optimize Google Gemini agents with Grafana Cloud

Manual prompt-by-prompt debugging fails when AI agents move to production environments serving hundreds or thousands of concurrent users. The Grafana Sigil SDK provides OpenTelemetry-native instrumentation that captures Gemini agent tool calls, execution decisions, and response latency without altering underlying business logic. Grafana Cloud AI Observability aggregates metrics including error rates, token consumption breakdown, cost estimations, and latency percentiles across model versions. Built-in AI evaluators and automated Slack assistants can diagnose operational bottlenecks such as rate-limit exhaustion and excessive token usage. Integration with Claude Code inside Slack allows engineering teams to close the remediation loop by generating and applying fixes to agent prompts and source code.

Sep 3, 2026Read summary
Thumbnail for End-to-end agentic development with Google Gemini, Antigravity, and GitLab8:13

End-to-end agentic development with Google Gemini, Antigravity, and GitLab

GitLab integrates directly with Google Cloud IAM and Google Artifact Registry to manage permissions and container images. Developers can use Gemini inside the Antigravity IDE to create visual UI mockups and translate design images into working application code. Model Context Protocol (MCP) server integration allows the Antigravity IDE to create and update GitLab issues and track planning workflows smoothly. GitLab Duo Agent Platform enables automated code review using specialized agents such as Develop with Gemini before merging code. GitLab CI/CD pipelines run automated security tests, upload build artifacts to Google Artifact Registry, and deploy directly to Cloud Run.

Sep 2, 2026Read summary
Thumbnail for He built a voice controlled drone to avoid using a remote!2:03

He built a voice controlled drone to avoid using a remote!

Bryen Param won the Live Agents category of the Gemini Live Agent Challenge with his drone-copilot project. The project replaces physical drone remote controllers with real-time voice commands. The backend infrastructure for drone-copilot is hosted on Google Cloud Run. The system uses the Gemini Live API for conversational flight control and Gemini Flash for visual inspection tasks. Details of the winning project and other competition entries are hosted on Devpost.

Sep 2, 2026Read summary
Thumbnail for Automate Google Cloud with Cloud CLI Remote MCP Server2:12

Automate Google Cloud with Cloud CLI Remote MCP Server

The Google Cloud CLI Remote MCP Server allows AI agents to interact with Google Cloud environments without installing local CLI binaries. Enabling the plugin equips an agent with tools to execute both gcloud and BigQuery (bq) commands through natural language prompts. During an incident triage demonstration, the agent queried Cloud Logging to locate a missing data file and inspected BigQuery tables to verify pipeline impact. Remediation actions that modify data require user confirmation before executing commands such as gcloud storage cp. All agent operations run under the user's IAM credentials within a secure sandbox to preserve enterprise governance.

Aug 28, 2026Read summary
Thumbnail for Launch your portfolio website with AI8:30

Launch your portfolio website with AI

Google AI Studio Build Mode allows developers to rapidly prototype full-stack and frontend web applications using natural language prompts. The build interface automatically generates code, provides selectable visual design themes, and displays a live interactive preview sandbox. Generative AI outputs are non-deterministic, meaning identical prompts can yield different code, layouts, and simulated background details. Developers can iterate on generated applications by asking the assistant in natural language to add, modify, or remove specific UI sections and features. Web applications created in Google AI Studio can be deployed directly to Google Cloud Run with a hosted URL ending in ai.studio or mapped to a custom domain. Completing the builder lab awards the Completed Builder Journey learning badge on Google Developer profiles.

Aug 28, 2026Read summary
Thumbnail for Launch your to-do web app with AI7:57

Launch your to-do web app with AI

A full-stack application consists of a front end user interface, a back end server connection, and a database for persistent storage. Google AI Studio Build Mode generates full-stack web applications from natural language prompts using Gemini models. Firebase Firestore integration allows applications generated in Google AI Studio to store data with real-time synchronization. Firebase Authentication provides built-in Google Sign-In support to restrict data access to specific users. Applications can be deployed to Cloud Run with a single click and given custom vanity URLs under the ai.studio domain.

Aug 28, 2026Read summary
Thumbnail for Launch your web-based video game with AI11:39

Launch your web-based video game with AI

Event sourcing records every game turn as a distinct sequential event rather than overwriting existing database entries. Google AI Studio Build Mode generates the full-stack web application, Firestore integration, and Minimax AI opponents from natural language prompts. Firestore security permission errors can be diagnosed and resolved automatically using the Fix feature in Google AI Studio. Firebase Firestore real-time data synchronization allows live multiplayer matches across separate devices without manual socket configuration. Logging event history enables time-travel features such as scrubbing through past turns and rolling back moves.

Aug 28, 2026Read summary
Thumbnail for 7 AI agent patterns to improve your coding workflow11:29

7 AI agent patterns to improve your coding workflow

Modular agent skills replace long, confusing prompts by allowing coding agents to read short skill descriptions and dynamically load full instructions only when needed. Skill files are plain text files stored either per project in the project root directory or globally in Gemini configuration directories. The seven core skill design patterns categorize capabilities into domain knowledge, tool wrappers, inversion, generators, reviewers, pipelines, and meta skills. Inversion patterns force the AI agent to stop guessing and ask clarifying questions or follow structured root cause analysis before writing code. Pipeline skills orchestrate sequential development workflows such as test-driven development through strict multi-step execution guards. Developers can create custom skills as text files following guided resources like Google codelabs and evaluate them using published whitepapers.

Aug 27, 2026Read summary
Thumbnail for What "real-time voice AI" actually means0:52

What "real-time voice AI" actually means

Conventional voice AI typically stitches together three separate systems comprising speech-to-text, a central model, and text-to-speech. Stitched voice pipelines force a turn-taking conversation dynamic where the system cannot hear the user while speaking. A real-time voice agent processes audio input and output simultaneously within a single live session. True real-time voice architecture supports barge-in, allowing a user to interrupt the agent mid-sentence and have it stop speaking immediately. Google Cloud demonstrates building genuine real-time conversational agents using Gemini Live.

Aug 27, 2026Read summary
Thumbnail for Build a real-time voice AI agent with Gemini Live API7:13

Build a real-time voice AI agent with Gemini Live API

Traditional text-to-speech AI systems operate in one direction, whereas Gemini Live provides bidirectional, simultaneous audio-to-audio streaming. The application architecture consists of three components: a browser frontend, a lightweight backend relay, and the Gemini Live API connected via WebSockets. The core interaction loop follows four simple steps: open the session, send microphone audio, receive model audio, and play the response in the browser. Voice Activity Detection (VAD) is built directly into Gemini Live to detect when speech starts and stops without requiring custom detection code. Barge-in allows users to interrupt the agent mid-sentence, which feels instantaneous when local audio playback is stopped immediately upon detecting mic input. External tools and functions allow the voice agent to execute real actions like controlling music playback, provided the functions return responses quickly.

Aug 26, 2026Read summary
Thumbnail for From notebook to production: Serving JAX at scale3:05

From notebook to production: Serving JAX at scale

Moving a trained JAX model from a notebook into production requires choosing a serving pattern that avoids unexpected just-in-time compilation latency during live user requests. Using jax.jit provides the simplest in-process Python serving path, but models should be warmed up with dummy inputs to absorb initial compilation costs before handling live traffic. Ahead-of-time compilation explicitly lowers Python code to StableHLO intermediate representation and compiles it to a device executable to eliminate first-call latency spikes. Portable deployments can be achieved using jax.export for serialized StableHLO artifacts or jax2tf to export standard TensorFlow SavedModel formats for existing TensorFlow serving infrastructure. Debugging and optimizing JAX on GPUs relies on a four-layer mental model consisting of JAX Python tracing, XLA compilation, CUDA or cuDNN execution, and GPU hardware.

Aug 26, 2026Read summary
Thumbnail for Scale JAX models to multi-GPU systems3:28

Scale JAX models to multi-GPU systems

Scaling JAX workloads across multiple GPUs focuses on array placement and sharding declarations rather than rewriting training routines with low-level communication code. Data parallelism in JAX operates by replicating model parameters across devices, splitting data batches into shards, and performing automatic gradient averaging across GPUs. Four core concepts structure multi-GPU placement in JAX: Mesh, PartitionSpec, NamedSharding, and jax.device_put. JAX supports both declarative automatic sharding and explicit low-level control through shard_map combined with jax.lax.pmean. A Tiny Transformer language model built with Flax NNX demonstrates multi-GPU training, checkpointing with Orbax, and length-padded JIT generation.

Aug 26, 2026Read summary
Thumbnail for Build and optimize JAX training loops3:25

Build and optimize JAX training loops

A clean JAX training step is designed as a pure function that ingests parameters, optimizer state, and a batch, returning updated states and metrics. Using fixed batch sizes enables JAX and XLA to reuse compiled execution steps on the GPU without triggering repeated recompilations. Converting GPU metrics to Python floats on every single step forces CPU-GPU synchronization and severely degrades training loop throughput. Naive attention materializes the full quadratic attention matrix, whereas fused implementations like cuDNN scaled dot-product attention drastically reduce memory and latency. Decoder architectures rely on causal attention masking, where Multi-Head, Grouped-Query, and Multi-Query Attention offer trade-offs in KV cache footprint while keeping the same output shape.

Aug 26, 2026Read summary
Thumbnail for Getting started with JAX on NVIDIA GPUs4:00

Getting started with JAX on NVIDIA GPUs

Verifying GPU execution requires checking hardware visibility with nvidia-smi as well as inspecting jax.devices and jax.default_backend in code. JAX distinguishes between host CPU memory and device GPU memory, meaning operations like converting to NumPy arrays or casting tensors to Python floats can cause synchronization bottlenecks in hot loops. The core functional transformations in JAX are jit for compiling, grad for automatic differentiation, and vmap for vectorization and batching. JAX compiles functions via XLA based on input shape and data type signatures, causing re-compilation whenever array shapes change dynamically unless padded with masks. Python conditional branching fails on traced values inside jit-compiled functions, requiring JAX-compatible control flow primitives like jnp.where, lax.cond, and lax.scan. Accurate benchmarking on GPUs requires calling block_until_ready on array outputs because JAX dispatches GPU kernels asynchronously.

Aug 26, 2026Read summary
Thumbnail for 4 ways loop engineering fails (and how to fix them)4:38

4 ways loop engineering fails (and how to fix them)

Loop engineering automates iterative prompting by running continuous retries until a specified goal is met. Runaway loops consume API tokens and generate financial costs unless restricted by iteration caps, time limits, or budget limits. Having a single agent review its own work causes confirmation bias and context pollution, which can be fixed by using deterministic checks or independent evaluator agents. AI loop objectives require explicit, non-negotiable criteria such as strict word count limits, zero compilation errors, or passing unit tests. When task complexity exceeds the capacity of a single loop, developers should transition to graph engineering to orchestrate multiple specialized agents.

Aug 25, 2026Read summary
Thumbnail for Building apps on Google Cloud1:06

Building apps on Google Cloud

Anyone with an app or website idea is considered a builder regardless of prior coding experience. Google provides a dedicated starting portal at goo.gle/builders containing bite-sized tutorials and labs. Available learning modules include launching AI-driven portfolio websites, to-do web apps, multiplayer games, and Android applications. The builder library is continuously expanding with new video walkthroughs and technical labs. Viewers are invited to request specific tutorials or share their completed projects through the video comments.

Aug 24, 2026Read summary
Thumbnail for How Globo leveraged Google Skills to certify 140+ AI professionals0:55

How Globo leveraged Google Skills to certify 140+ AI professionals

Artificial intelligence is an established core component of Globo's corporate strategy. Globo links employee training directly to measurable initiatives and product integration. Google Skills provides Globo with structured training programs to advance internal AI competencies. Around 500 Globo employees are currently registered on the Google Skills platform. More than 140 Globo team members have earned official Google certifications.

Aug 21, 2026Read summary
Thumbnail for New Way Now: Personal builds the agentic AI Concierge with Google Cloud Consulting2:26

New Way Now: Personal builds the agentic AI Concierge with Google Cloud Consulting

Personal developed Concierge, an agentic artificial intelligence platform powered by generative AI that autonomously resolves complex customer inquiries. The Concierge platform has processed over two million customer interactions with an effectiveness rate exceeding 87 percent in identifying interactions. The solution integrates Google Cloud products including Customer Engagement Suite, Contact Center AI, and Gemini models via Vertex AI. Personal partnered with Globant for agile execution and development of use cases, while Google Cloud Consulting provided architectural guidance and technical support. The system uses custom voice technology and natural language processing to deliver more agile and natural customer service across channels.

Aug 21, 2026Read summary
Thumbnail for Prism Distributed Inference Performance4:16

Prism Distributed Inference Performance

Prism is a performance analysis tool for llm-d designed to help teams configure distributed inference stacks using ground-truth benchmarks. The tool organizes distributed inference patterns into foundational building blocks, such as intelligent routing and cache offloading, as well as integrative workloads like agentic serving. In intelligent routing benchmarks, approximate prefix cache-aware routing achieved a 98 percent latency reduction and a 147 percent throughput increase over standard round-robin Kubernetes routing. For agentic code generation workloads, combining cache-aware routing with kernel optimizations yielded a 111.6 percent throughput increase and a 92.4 percent latency reduction. Prism supports multiple engineering roles across the full benchmarking lifecycle, incorporating real-world workload catalogs, test harnesses, and an open results store.

Aug 21, 2026Read summary
Thumbnail for LLM serving with SGLang TPU4:08

LLM serving with SGLang TPU

SGLang-JAX is built from the ground up to serve large language models on Google Cloud TPUs using custom Pallas kernels and a JAX backend. The framework supports modern state of the art models including Qwen, DeepSeek, GLM, and Xiaomi MiMo-v2-Flash. Setting up and serving models with SGLang on TPU environments matches the standard workflow used on GPU platforms. Benchmark tests demonstrate that serving performance on TPUs achieves parity with GPU deployments. Google partnered with the open source SGLang community to bring native TPU support and performance optimizations to the framework.

Aug 21, 2026Read summary
Thumbnail for Embedding models with vLLM TPU4:03

Embedding models with vLLM TPU

Google Cloud has integrated Cloud TPU support into vLLM to serve embedding models alongside large language models. Embedding models convert input data such as text into dense mathematical vectors for semantic search, recommendation engines, and intent classification. Serving embedding models with vLLM on TPUs allows infrastructure teams to standardize their serving framework and scale across both GPU and TPU nodes. The vLLM framework optimizes embedding model weights across TPU cores using underlying XLA compilation. A demonstration on a Cloud TPU v6e VM shows that vLLM exposes an OpenAI-compatible HTTP endpoint for embedding generation and supports high-throughput benchmarking.

Aug 21, 2026Read summary
Thumbnail for GPU Reliability5:20

GPU Reliability

Google Cloud partners with Nvidia during the product initiation phase to build reliability into GPU hardware before deployment. Before delivering GPU capacity to customers, Google runs foundational stress tests and uses the Nvmitron framework to produce a bill of health report. Post-deployment observability relies on metrics such as mean time between interruption and goodput to assess infrastructure utilization and stability. Google Cloud provides preconfigured observability dashboards in Cluster Director Health Monitoring to track streaming multiprocessor utilization, GPU power, thermal margins, and temperatures. The Emergent Maintenance feature alerts customers up to seven days in advance when anomalies like thermal spikes are detected, allowing scheduled repairs without abrupt disruption.

Aug 21, 2026Read summary
Thumbnail for Compute Infra4:19

Compute Infra

Google Cloud Compute Engine offers workload-optimized virtual machine families divided into general-purpose and specialized categories. General-purpose N-series machines support custom machine types that can reduce compute spend by up to 19 percent. Specialized M-series and X-series machines deliver a 99.95 percent single-instance service level agreement for mission-critical databases. Network-optimized C4N instances provide sustained packet throughput of up to 95 million packets per second for network appliances and telecommunications workloads. Google Cloud Hyperdisk allows independent tuning of storage capacity and performance across Balanced, Throughput, Extreme, and ML options.

Aug 21, 2026Read summary
Thumbnail for TPU Reliability2:45

TPU Reliability

Training frontier large language models with trillions of parameters transforms training into a distributed systems problem vulnerable to hardware failures and network latency. Google Cloud TPU 7x Ironwood systems deploy at a pod scale of 9,216 chips and 1.77 petabytes of high-bandwidth memory across 144 cubes. Hardware fault tolerance is handled through rapid cube-level hot swaps from a holdback pool within seconds rather than troubleshooting individual machines for minutes. Interconnect technologies like ICI, 3D Torus, and Optical Circuit Switches enable dynamic reconfiguration and connectivity across thousands of TPU chips. Software resiliency pairs framework-level management with application-level automated checkpointing, failure interception, and rollback recovery to minimize training downtime.

Aug 21, 2026Read summary
Thumbnail for Flexible Compute3:52

Flexible Compute

Google Cloud AI Hypercomputer includes a flexible consumption layer featuring the Dynamic Workload Scheduler, also known as DWS. Dynamic Workload Scheduler addresses on-demand capacity predictability risks by balancing predictability, commitment, and pricing for short-term workloads. DWS Flex mode provisions capacity as it becomes available for workloads with flexible execution timelines, supporting run durations of up to seven days. DWS Calendar mode allows users to reserve required compute capacity for specific future dates and times up to 90 days in advance. Using DWS Flex mode requires an existing Google Kubernetes Engine cluster to create node pools configured with the flex start flag.

Aug 21, 2026Read summary
Thumbnail for Intro to AI Hypercomputer3:28

Intro to AI Hypercomputer

AI infrastructure must support the entire model lifecycle, including training, fine-tuning, serving, and reinforcement learning. Scaling compute does not yield linear performance gains due to bottlenecks in node placement, networking, storage, and checkpointing. Google Cloud AI Hypercomputer addresses these challenges through a co-designed architecture spanning hardware, software, and consumption layers. The hardware foundation combines custom TPUs, GPUs, and CPUs with high-bandwidth networking and AI-optimized storage. Open software frameworks, compilers like XLA, and orchestrators like GKE and Dynamic Workload Scheduler enable flexible and efficient workload execution.

Aug 21, 2026Read summary
Thumbnail for This or that? Imaxxing edition0:53

This or that? Imaxxing edition

Andrew Baker chooses spaces over tabs, noting that mixing the two creates problems. Baker prefers uv to pip for Python package management due to its speed. Between consistency and availability, Baker prioritizes availability, pointing to Temporal's uptime record displayed on its homepage. Baker selects orchestration over choreography for managing workflows. For lifestyle choices, Baker prefers candy like Milk Duds over popcorn, window seats over aisle seats, and the movie Inception over The Prestige.

Aug 20, 2026Read summary
Thumbnail for Google Skills empowers Palo Alto Networks with 24/7 on-demand training0:42

Google Skills empowers Palo Alto Networks with 24/7 on-demand training

Palo Alto Networks provides employees with hands-on opportunities to build models and experience technical tools directly. Google Skills serves as an ongoing reference point and reinforcement tool as employees develop their technical expertise. The 24/7 availability of self-paced learning accommodates the company's globally distributed workforce. A centralized learning platform ensures consistent educational and onboarding opportunities across all teams.

Aug 20, 2026Read summary
Thumbnail for How Google Skills drove EVERSANA AI adoption from 3% to 70%0:39

How Google Skills drove EVERSANA AI adoption from 3% to 70%

EVERSANA set out an initiative to train all of its employees to use artificial intelligence tools. Initial internal adoption was low, with AI usage hovering between 3 percent and 5 percent during the first few weeks. Partnering with Google enabled EVERSANA to implement structured learning and instructor-led training. Following the Google Skills training program, employee AI adoption increased from 3 percent to 70 percent.

Aug 20, 2026Read summary
Thumbnail for Google Accelerator Alumni: Tilli1:37

Google Accelerator Alumni: Tilli

Tilli is an early education startup with the mission of ensuring every child is developmentally on track by their tenth birthday. During the Google Accelerator program, Tilli achieved its target of securing its first 50 paying schools. Tilli developed a mobile platform enabling teachers to query student developmental data and receive personalized four-week intervention plans. The platform has tracked data across five years to support 36,000 children in mastering 12 foundational skills at grade level. Using Google AI Studio for rapid prototyping reduced Tilli's software development cycle from ten weeks to a single day. Tilli is expanding its educational footprint from 50 to 100 schools across South Asia and the Middle East.

Aug 20, 2026Read summary
Thumbnail for How to modernize legacy codebases using AI coding agents (Step by step guide)8:25

How to modernize legacy codebases using AI coding agents (Step by step guide)

AI coding agents can trace data dependencies across frontends, APIs, and databases to update complex legacy codebases. A change request driven strategy breaks legacy modernization into three distinct phases: plan, execute, and verify. Documenting local toolchains in a README file allows AI agents to run compilers, linters, and tests to autonomously self-correct syntax errors. Planning mode enables agents to map the entire blast radius of a change across files and database tables before editing code. Decoupling test authoring from code generation prevents AI agents from creating self-serving tests that mask implementation flaws. A three-layer verification system consisting of human review, secondary CI/CD AI review, and staging smoke tests ensures bugs do not reach production.

Aug 20, 2026Read summary
Thumbnail for New Way Now: KDDI pioneers Japan's AI ecosystem with Google Cloud Consulting1:41

New Way Now: KDDI pioneers Japan's AI ecosystem with Google Cloud Consulting

KDDI is a major Japanese telecommunications carrier offering consumer services such as financial services, cable television, and convenience stores, alongside enterprise solutions like data centers and cloud integration. With domestic telecommunications market growth slowing down in Japan, KDDI is focusing on artificial intelligence to develop new revenue streams beyond traditional connectivity. Google Cloud Consulting and its Black Belt team partnered with KDDI to provide hands-on technical guidance and architecture expertise. The collaboration assisted KDDI in establishing technical architecture and determining how to build new AI-driven service offerings.

Aug 19, 2026Read summary
Thumbnail for Real-time voice AI agents, explained (before you build one)6:26

Real-time voice AI agents, explained (before you build one)

Omni apps represent a new software paradigm that processes multimodal inputs and outputs continuously rather than through turn-based text prompts. A bidirectional live loop allows users and AI agents to communicate simultaneously over a single open connection, enabling natural interruptions mid-sentence. Google's Agent Development Kit and the GenAI SDK provide frameworks to manage the real-time flow of audio, events, and tool invocations. Tool call latency presents a critical challenge in live voice interfaces, where delays of a few seconds can disrupt conversational flow. A full-stack live multimodal agent requires integrating speech, framework connectivity, tool calling, browser automation, hybrid memory retrieval, and camera vision.

Aug 19, 2026Read summary
Thumbnail for Preview: building a viral movie app with Temporal & Cloud Run0:42

Preview: building a viral movie app with Temporal & Cloud Run

IMAXXING is a web application created to monitor movie showtimes across the country and alert users when seats open up. The application was inspired by high demand and sold-out 70mm IMAX tickets for The Odyssey. The system architecture runs on Google Cloud and Temporal. The service is deployed on Google Cloud Run using pre-release Temporal Serverless Workers.

Aug 19, 2026Read summary
Thumbnail for Building a viral Odyssey movie app with Temporal & Cloud Run11:41

Building a viral Odyssey movie app with Temporal & Cloud Run

IMAXxing is an automated notification service built to monitor 70mm IMAX seat availability across the country for Christopher Nolan's The Odyssey. The application architecture uses Temporal for durable workflow orchestration and Google Cloud Run for running serverless worker capacity. Temporal workflows track user subscriptions as individual entity workflows and use signal events and timers to batch alerts and prevent excessive notifications. Using serverless workers with Temporal allows the system to scale compute resources directly according to task queue demand during sudden booking spikes. AI coding agents and tools like Terraform and the gcloud CLI significantly accelerated the infrastructure deployment and monitoring configuration.

Aug 18, 2026Read summary
Thumbnail for New Way Now: Ericsson boosts ROI and issue resolution with Gemini Enterprise Agent Platform1:46

New Way Now: Ericsson boosts ROI and issue resolution with Gemini Enterprise Agent Platform

Operations support systems manage service activation while business support systems handle billing and monetization. Ericsson collaborates with Google Cloud to simplify telecommunications operations and convert emerging technologies into business value. Google's Gemini Enterprise Agent Platform aggregates network data to power autonomous agents capable of closed-loop issue detection and remediation. Using Google Cloud AI tools allows telecom operators to reduce root cause analysis turnaround times from hours down to minutes.

Aug 18, 2026Read summary
Thumbnail for Cloud Network Insights Cloud Network Insights: end-to-end observability for the Cross-Cloud Network18:57

Cloud Network Insights Cloud Network Insights: end-to-end observability for the Cross-Cloud Network

Cloud Network Insights is a fully managed, AI-powered network, web application, and agent performance monitoring tool designed for distributed, multi-cloud, and hybrid environments. The product offers visibility across Google Cloud, third-party clouds, on-premises data centers, internet-facing services, APIs, and transit networks. Cloud Network Insights integrates with Google Cloud observability tools like connectivity tests and VPC flow logs, and supports OpenTelemetry Protocol for metrics and logs. Pricing is based on active paths being monitored, with pay-as-you-go, enterprise, and unlimited licensing options available. The monitoring capabilities cover network performance metrics such as latency, loss, and jitter, as well as layer 7 web application monitoring and milestone scripting.

Aug 14, 2026Read summary
Thumbnail for Steal this Agent0:51

Steal this Agent

Laura Mae Martin introduces Gemini Enterprise as a tool to automate weekly progress reports. Users can build specialized AI teammates in plain English without coding. The agent builder requires defining a role, linking workspace sources, and specifying the output format. Gemini Enterprise secures data by querying only accessible files and keeping company data within the domain.

Aug 14, 2026Read summary
Thumbnail for New Way Now: PayPal paving the way for the future of agentic commerce with Google Cloud3:23

New Way Now: PayPal paving the way for the future of agentic commerce with Google Cloud

PayPal serves millions of consumers and processes $1.8 trillion annually, making trust and scale critical priorities. The company migrated its entire AI stack to Google Cloud to gain compute horsepower and scaling flexibility. Transitioning to Google Kubernetes Engine reduced release lifecycles from weeks of regression testing down to minutes. PayPal supports over 500 models in production and handles nearly one billion requests every single day. Using MCP and AP2 protocols, PayPal enables autonomous agents to interact safely and audibly with payment systems.

Aug 13, 2026Read summary