googleblog.com

50 articles in the TruthFoundry index. Each links out to the original.

• Scaling Real-Time AI Agents with Session-Aware Load Balancing · Developers must combine CPU metrics with active session counts to balance real-time AI agent workloads effectively.
• Ray AI Libraries Enable Distributed Training and Serving on Google TPUs · Ray Serve, Ray Data, and Ray Train now support distributed AI workloads on Google Cloud TPUs with automatic slice placement.
• Google Announces LiteRT.js for High-Performance Web AI Inference · Google releases LiteRT.js, enabling native, hardware-accelerated AI models in web browsers via WebAssembly.
• Google ADK 2.0 Adds Deterministic Workflows for Enterprise AI Agents · Google ADK 2.0 introduces deterministic workflows to bridge the gap between flexible AI agents and strict business logic.
• Google Developers Build AI Race Coach Using Antigravity and Gemini · Google Developer Experts deployed an AI race coach at Sonoma using Antigravity and Gemini to provide real-time driving advice.
• Conductor Evolves into Plugin Supporting Antigravity CLI · Conductor transitions to a plugin architecture to support Antigravity CLI and enhance conversational spec-driven development.
• Google Genkit Adds Agent Skills for Modular AI Workflows · Firebase Genkit introduces Agent Skills to modularize AI agent expertise and reduce prompt context costs.
• Firebase Genkit Adds Agents API for Full-Stack AI Apps · Firebase Genkit launches a preview Agents API to simplify building conversational AI applications.
• Google Announces GA for Agent and Model Evaluations in Gemini Enterprise · Google has made agent and model evaluation tools generally available for the Gemini Enterprise Agent Platform.
• ADK Go 2.0 Introduces Graph-Based Workflow Engine for Multi-Agent Apps · Google's ADK Go 2.0 launches a graph-based workflow engine with built-in human-in-the-loop and dynamic orchestration.
• Google Releases Coding Agent Skill to Drive Agent Quality Flywheel · Google launches a developer skill to automate testing and improving AI agents using adaptive evaluation.
• Google Microbenchmarks Guide Evaluates TPU Performance · Google's microbenchmark suite evaluates TPU performance across network, compute, and memory metrics to optimize hardware efficiency.
• Google ADK Enables Cross-Language Python and Go Agent Collaboration via A2A · Google's Agent Development Kit allows Python and Go agents to collaborate using the A2A protocol for contract compliance.
• Ironwood Platform Achieves 3.1x-4.7x Speedup for Qwen 3.5 MoE Inference · Google engineers optimized Qwen 3.5-397B on Ironwood TPUs, achieving 3.1x decode and 4.7x prefill speedups.
• Google Labs Tests Proactive AI Coding Agents on Real Bug Clusters · Google Labs evaluates proactive AI coding agents using real bug clusters to measure diagnostic insight.
• Google Releases Tunix for High-Throughput Agentic RL Training · Google's Tunix library solves infrastructure bottlenecks in training reasoning agents with asynchronous rollouts and composable environments.
• Google's MaxText Enables Elastic TPU Training Recovery in Seconds · MaxText and PathwayS allow TPU training jobs to recover from worker failures without restarting the entire process.
• Ray 2.55 Adds Official TPU Support via Google Kubernetes Engine · Ray 2.55 introduces official TPU support, allowing Python code to run on Google Cloud TPUs using standard APIs.
• Google's A2A Protocol Celebrates One Year as Agents Collaborate Securely · Google's Agent-to-Agent protocol enables autonomous AI collaboration, highlighted by FoldRun's protein modeling capabilities.
• Google Announces Agentic Resource Discovery Specification for AI Tools · Google launches open specification for discovering and verifying AI agents and tools across the web.
• New Architectural Patterns Merge A2UI and MCP for Unified Agent Interfaces · Developers can now combine A2UI's native rendering with MCP's connectivity to solve UI fragmentation.
• Google Partners with Parallel Web Systems for Gemini Enterprise Agent Platform · Google Cloud integrates Parallel Web Systems as a native grounding provider for Gemini Enterprise Agent Platform.
• Building Scalable AI Agents with Modular Prompt Transpilation · Production AI agents require modular prompt transpilation to ensure reliability and maintainability.
• Google Launches TPU Developer Hub for AI Model Builders · Google introduces a new Developer Hub to help users optimize and deploy models on TPU infrastructure.
• Google Releases DiffusionGemma Developer Guide for Experimental AI Model · Google shares a developer guide for DiffusionGemma, an experimental non-autoregressive model built on the Gemma 4 backbone.
• Google Announces Colab CLI for Agent-Driven Machine Learning Workflows · Google introduces a Command-Line Interface for Colab to enable seamless remote execution for developers and AI agents.
• Google Unveils Gemma 4 12B for Local Laptop AI Workflows · Google DeepMind releases Gemma 4 12B model optimized for local execution on everyday laptops via Google AI Edge.
• Google Unveils Gemma 4 12B: Developer-Friendly Multimodal AI Model · Google releases Gemma 4 12B, a developer-friendly multimodal AI model with an encoder-free architecture and local GPU support.
• Google Adds New Session Metadata Claims to Sign in with Google · Google introduces auth_time and amr claims to enhance security and trust in Sign in with Google.
• Google Cloud Workbench Extension for VS Code Now Available to Enhance ML Development · Google Cloud Workbench extension for VS Code launches, enabling seamless ML development with cloud infrastructure