Google I/O 2026 Pushes Agentic Gemini Across Apps and APIs

What shipped

On 19 May, Google I/O delivered a broad set of agentic updates to Gemini models, apps, and APIs alongside infrastructure moves from NVIDIA and Vercel. The releases focus on turning prompts into actions that run in sandboxes or on existing cloud setups. Builders now face more choices for shipping proactive assistants without starting from scratch.

Vendor launches

Google announced over twenty updates centered on agent capabilities in Gemini 3.5, managed agents, and new workspace tools. NVIDIA highlighted cost reductions for inference and faster sandboxes on its Vera hardware used by thousands of enterprises. Vercel added flat-rate pricing and Claude agent support to make production runs more predictable.

• •NVIDIA Google Cloud Community The joint program now supports over 100,000 developers with labs and events for building on the full NVIDIA stack inside Google Cloud. Vibe builders gain ready paths to test agent workloads without managing hardware.
• •Google Flow Upgrades Google Flow and Flow Music gained new agents, mobile apps, and Gemini Omni for turning any input into editable creative output. Non-techies can generate music or visuals through simple conversation.
• •Gemini App Agentic Mode The Gemini app now runs proactive 24/7 assistance that anticipates needs across daily tasks. SMB owners can delegate routine follow-ups without constant prompting.
• •Gemini for Science New experiments and tools let researchers scale precision in discovery workflows. Developers can plug these into existing lab pipelines for faster hypothesis testing.
• •Google AI Ultra Plan A new $100 tier adds premium benefits across AI subscriptions for heavy users. Teams get higher limits when running multiple agents daily.
• •AI Search Evolution Search now blends classic results with AI summaries that handle natural language questions. Users shift from keywords to full sentences for quicker answers.
• •Universal Cart An intelligent shopping cart gathers items across sites and lets agents complete purchases. SMB owners can test automated buying flows for repeat orders.
• •Intelligent Eyewear Gemini-powered glasses arriving this fall handle directions, texts, and photos hands-free. Field teams gain real-time assistance without pulling out phones.
• •Managed Agents in Gemini API Developers can define agents as files and run them in secure cloud sandboxes. Production teams reduce setup time when deploying task-specific agents.
• •Gemini 3.5 Release The new model series adds action-taking ability to frontier-level reasoning. Builders can run parallel agent loops for complex multi-step jobs.
• •Agentic Gemini Era Sundar Pichai framed the conference around getting more done with proactive agents. Companies see faster internal workflows once agents handle routine coordination.
• •Gemini Omni The model creates or edits content from any input using conversational edits. Creators iterate designs without switching tools.
• •Workspace Voice Updates Gmail, Docs, and Keep now accept voice commands plus a new design tool called Google Pics. Teams draft and organize work faster during meetings.
• •AI Mode Search Trends After one year, users in the US prefer natural language queries over keywords. Search volume shows clearer intent signals for marketers.
• •I/O 2026 Overview Google collected all announcements in one hub covering agents, models, and hardware. Builders can scan the full list to pick integrations quickly.
• •Android Halo Status bar integration surfaces agent intelligence directly on Android devices. Mobile users receive live updates without opening separate apps.
• •Stitch Design Agent The tool turns live collaboration with an agent into real-time design changes. Vibe builders prototype interfaces through conversation instead of manual edits.
• •Developer Highlights Updates to Google Antigravity and the Gemini API move ideas from prompt to production apps. Engineers benchmark these against current stacks before migration.
• •Pomelli Brand Agents New agents build brand books and launch websites from brief inputs. Small businesses generate consistent marketing assets in hours instead of days.
• •Gemini 3.5 Flash on Vercel The model now runs on AI Gateway with better coding and parallel agent execution. Teams cut latency on routine tasks while keeping costs lower than prior versions.
• •Flat Rate CDN Beta Pro teams on Vercel can switch to fixed monthly CDN pricing to avoid traffic spikes. Cost predictability improves when running high-volume agent endpoints.
• •Claude Agents on Vercel Sandbox Claude Managed Agents now execute inside isolated Firecracker microVMs on existing Vercel infrastructure. Developers keep private APIs and customer data inside their current environment.
• •Gemini Education Studies Trials in Sierra Leone and Italy showed math skill gains for students and faster lesson prep for teachers. Schools can pilot similar setups with measurable outcome tracking.
• •NVIDIA Vera at Dell World Jensen Huang reported parabolic demand and claimed agent sandboxes run 50 percent faster on Vera than CPUs. Enterprises like Lilly and Samsung already run workloads on Dell AI Factories.

Product Hunt picks

Five new tools appeared on Product Hunt focused on voice agents, training data, and verification. Cursor shipped a stronger model while niche entries target language coverage and authenticity checks.

• •Composer 2.5 Cursor released its strongest coding model yet for complex refactoring tasks. Vibe builders can tackle larger codebases with fewer manual corrections.
• •Thinnest AI The service builds voice agents in over 100 languages at 1.5 rupees per minute. SMB owners test multilingual support bots without large upfront spend.
• •VWFNDR + MBL The camera app captures raw photos with cryptographic proof they are not AI-generated. Creators protect original work when sharing on public platforms.
• •Trainer Users record their screen to train custom AI agents for repeated desktop tasks. Non-coders automate workflows by demonstration instead of scripting.
• •Odysseyml The platform runs multiplayer game simulations inside an AI world model. Developers prototype game mechanics faster before full engine integration.

Replicate new models

HeyGen Avatar V: The engine produces realistic talking avatar videos from text prompts using cross-reference animation. Vibe builders embed short clips into marketing sites without video crews.

Hugging Face trending

Seventeen models and papers trended on Hugging Face covering unified multimodal systems, long-video infrastructure, and agent reasoning improvements. Labs released practical benchmarks for video, time-series, and early-exit inference.

• •HRM-Text-1B Sapient released a 1B text-generation model now trending for general fine-tuning. Builders test small-scale agents on local hardware.
• •Lance by Bytedance The any-to-any model handles image and video tasks through multi-task training. Teams compare it against larger unified models for editing speed.
• •LiteFrame Vision Encoders Research shows efficient encoders that scale video LLMs to longer clips by cutting token growth. Developers reduce memory use on extended footage.
• •STT-Arena Benchmark A new environment tests tool-using agents under spatio-temporal disruptions. Researchers measure replanning accuracy in realistic settings.
• •HINT-SD Method Hindsight self-distillation supplies intermediate hints for long-horizon agents. Training loops improve credit assignment on sparse rewards.
• •SNLP Inference Structured Newton corrections enable layer-parallel execution in autoregressive models. Latency drops on multi-layer stacks without accuracy loss.
• •KinGPT Chess Test A study probes whether fine-tuned models truly understand chess or simply memorize patterns. Results flag brittleness in game-playing claims.
• •LongLive-2.0 Infrastructure NVFP4 parallel setup speeds training and inference for long video generation. Studios cut wall-clock time on extended sequences.
• •Post-Trained MoE Skip Self-distillation lets mixture-of-experts models deactivate half their experts at inference. Compute savings appear on large language workloads.
• •SWIM Alignment The strategy aligns vision and language features for fine-grained video object understanding from text alone. Creators query scenes without drawing masks.
• •Semantic Early Exit Reasoning models stop once answers stabilize to avoid wasted tokens. Production pipelines gain lower latency on math and logic tasks.
• •Audio-Image Retrieval Onomatopoeic image matching helps locate sound effects for comics and media. Editors speed up asset searches with visual prompts.
• •Code as Agent Harness Papers explore using code itself as the runtime substrate for agent systems. Engineers embed tools directly inside generated scripts.
• •Lance Unified Model A lightweight multimodal model covers understanding, generation, and editing for images and video. Smaller teams run unified pipelines on modest GPUs.
• •AI Auto-Research Guide A roadmap outlines $15 paper generation and long-horizon experiment agents. Labs weigh integrity risks before scaling automation.
• •POST Anomaly Detection Prior-observation adversarial learning improves multivariate time-series anomaly detection. Monitoring systems catch spatial over-generalization earlier.
• •PySIFT GPU SIFT A deterministic SIFT implementation runs entirely on GPU for vision pipelines. Deep learning stacks gain classical feature stability without CPU round-trips.

Other

Global MMLU: Cohere released an expanded benchmark covering more languages and cultural contexts. Teams compare model performance beyond English-only tests.

Industry news

Six-Month LLM Recap: Simon Willison highlighted the November 2025 shift toward stronger coding agents. Builders track similar inflection points for planning next integrations.

What this means for you

What to watch next

Watch for production rollouts of Gemini 3.5 agents in Workspace and shopping. Track early benchmarks from teams running Claude agents on Vercel Sandbox. Note any follow-up releases from smaller labs that appeared on Hugging Face this week.

Sources

Vendor launches

• •NVIDIA and Google Cloud Empower the Next Wave of AI Builders
• •New agents, mobile apps and Gemini Omni for Google Flow and Google Flow Music
• •The Gemini app becomes more agentic, delivering proactive, 24/7 help
• •Gemini for Science: AI experiments and tools for a new era of discovery
• •Everything new in our Google AI subscriptions, fresh from I/O 2026
• •A new era for AI Search
• •Introducing the Universal Cart and more ways to help you shop
• •Intelligent eyewear is coming this fall
• •Introducing Managed Agents in the Gemini API
• •Gemini 3.5: frontier intelligence with action
• •I/O 2026: Welcome to the agentic Gemini era
• •Introducing Gemini Omni
• •New ways to create and get things done in Google Workspace
• •How AI Mode is changing the way people search in the U.S.
• •I/O 2026
• •Stay in sync with your agent with Android Halo.
• •We’re introducing new ways to design in real time with Stitch.
• •Building the agentic future: Developer highlights from I/O 2026
• •Pomelli adds new ways to build brand content and design websites.
• •Gemini 3.5 Flash on AI Gateway
• •Flat Rate CDN in Limited Beta
• •Run Claude Managed Agents with Vercel Sandbox
• •Measuring the impact of AI on teaching and learning
• •NVIDIA CEO Jensen Huang at Dell Technologies World: “Demand Is Going Parabolic, Utterly Parabolic”

Product Hunt picks

• •Composer 2.5
• •Thinnest AI
• •VWFNDR™ + MBL
• •Trainer
• •Odysseyml

Replicate new models

• •avatar-v by heygen launches on Replicate

Hugging Face trending

• •HRM-Text-1B by sapientinc trends on HuggingFace
• •Lance by bytedance-research trends on HuggingFace
• •LiteFrame: Efficient Vision Encoders Unlock Frame Scaling in Video LLMs
• •STT-Arena: A More Realistic Environment for Tool-Using with Spatio-Temporal Dynamics
• •HINT-SD: Targeted Hindsight Self-Distillation for Long-Horizon Agents
• •SNLP: Layer-Parallel Inference via Structured Newton Corrections
• •Generalization or Memorization? Brittleness Testing for Chess-Trained Language Models
• •LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation
• •Post-Trained MoE Can Skip Half Experts via Self-Distillation
• •See What I Mean: Aligning Vision and Language Representations for Video Fine-grained Object Understanding
• •Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models
• •Audio-Image Cross-Modal Retrieval with Onomatopoeic Images
• •Code as Agent Harness
• •Lance: Unified Multimodal Modeling by Multi-Task Synergy
• •AI for Auto-Research: Roadmap & User Guide
• •POST: Prior-Observation Adversarial Learning of Spatio-Temporal Associations for Multivariate Time Series Anomaly Detection
• •PySIFT: GPU-Resident Deterministic SIFT for Deep Learning Vision Pipelines

Other

• •Global MMLU

Industry news

• •Here’s why Elon Musk lost his suit against OpenAI
• •The last six months in LLMs in five minutes