Get tomorrow's brief in your inbox
Today: Google rolled out Gemini Spark, a cloud-based AI agent that runs 24/7 across your Google apps and external tools. Nvidia's new Vera CPU opened a $200 billion market for agentic AI (already $20B in sales). IrisGo launched desktop agents that learn your workflows and automate them without prompts.
Google just shipped its answer to OpenClaw: Gemini Spark, an AI agent that runs continuously in the cloud and connects to Gmail, Calendar, Drive, Docs, Sheets, and over 30 external services like Dropbox, Uber, and Spotify. Unlike OpenClaw's laptop-dependent approach, Spark runs 24/7 without keeping a device awake, syncing across web, Android, and iOS.
The big advantage here is Google's ecosystem lock-in. Where OpenClaw built adoption by integrating with apps people already use, Google can go deeper with native access to your email, calendar, and search history. It's entering trusted tester access this week, with a US beta next week for Ultra plan subscribers.
For small businesses, this means automation potential you couldn't get before without hiring someone to babysit Zapier workflows. Google's betting you'll use agents for shopping, research, and coordinating schedules, but the real value is in whatever custom workflows you discover. The downside: you're betting on Google's security model. Every task runs in an isolated VM with DLP policies, but you're still piping sensitive business data through an AI system that could be vulnerable to prompt injection attacks. (The Verge, Simon Willison)
IrisGo ($2.8M seed from Andrew Ng's AI Fund) launched desktop companions for Mac and Windows that learn your workflows by watching you do them once, then automate those tasks on command. During a demo, the founder showed it learning to order coffee online, then repeating that order autonomously. It includes a built-in skills library (email drafting, invoice processing, report building) plus learns from your desktop behavior to add new automation options. Privacy-conscious design processes most data on-device, with cloud processing only when you authorize it. Beta apps are live now, and IrisGo just struck a deal to pre-install on Acer laptops. If you run repetitive business processes (invoicing, report generation, data entry), this is worth testing. (TechCrunch)
Figma's AI agent launched in Figma Design with natural language controls for generating designs, editing existing work, and automating iterations. You can run multiple agents simultaneously on the canvas, and the company says models are fine-tuned specifically for design contexts. The agent will eventually roll out to other Figma products. If you prototype interfaces or mockups, this cuts the tedious parts of iteration work. (TechCrunch)
Google I/O announcements: Gemini 3.5 Flash (available now via API) delivers frontier intelligence at speeds comparable to the Flash series. It outperforms Gemini 3.1 Pro on coding and agentic benchmarks while costing less than half of other frontier models. Gemini Omni (video generation from any input) is rolling out to AI Plus, Pro, and Ultra subscribers globally. (Google AI Blog)
Ramp's engineering team is using OpenAI's Codex with GPT-5.5 to accelerate code review from hours to minutes and to build On-Call Assistant, an agentic tool that handles most on-call rotation work for engineers. The code review catches issues that human reviewers and other AI tools miss, and engineers specifically request it by name. The On-Call Assistant handles the complexity of incident investigations (concurrency bugs, external vs internal events, evolving details) that normally require intense mental effort and unbroken focus. If you manage a technical team, this is the use case: deploy AI to handle the grinding, context-heavy work that burns people out. (OpenAI)
n8n's guide to event-driven architecture explains how to implement async workflows without tight coupling between services. The article outlines real-world EDA scenarios (e-commerce checkout flows, fraud detection, inventory updates) and shows how n8n acts as an orchestration layer that manages event-driven workflows without sacrificing control or observability. If you're building or scaling any system that handles real-time events (order processing, notifications, data syncing), this is practical implementation guidance. (n8n Blog)
Google's Antigravity CLI (closed source, written in Go) is replacing the open source Gemini CLI on June 18th. If you've built workflows around the Gemini CLI, you have less than a month to migrate. The new CLI is bundled with Antigravity SDK (a Python wrapper around the closed source binary). (Simon Willison)
Gemini 3.5 Flash pricing: Less than half the cost of other frontier models with comparable performance. If you're running LLM workflows at scale (code generation, document processing, agent tasks), this cuts your API bill significantly. Available now via Gemini API in Google AI Studio and Android Studio. (Google AI Blog)
Nvidia reported $81.6 billion in revenue for the quarter ending April 26 (up 20% from previous quarter), with $75.2 billion in data center revenue. The company also revealed $43 billion in holdings in private startups, nearly double the $22 billion from January. Nvidia committed $30 billion to OpenAI in February and is scaling capacity for Anthropic this year. For SMBs, this signals continued infrastructure investment in AI (compute capacity isn't disappearing), but also shows how much capital is concentrating at the top of the stack. (TechCrunch)
Nvidia's Vera CPU is positioned as the world's first CPU built specifically for agentic AI. Jensen Huang says it unlocks a "brand new $200 billion TAM" and has already sold $20 billion worth this year. The logic: while model "thinking" uses GPUs, agents run tasks on CPUs. Vera is optimized for token processing speed rather than multi-core parallelism. If the prediction holds that billions of agents will eventually run like PCs, this is infrastructure investment paying off early. For businesses building or deploying agents, Vera becomes the hardware target. (TechCrunch)
AI search startups raised major rounds: Exa Labs raised $250M at a $2.2B valuation (Andreessen Horowitz), Parallel Web Systems raised $100M at $2B valuation (Sequoia), and a wave of startups (Tavily, TinyFish) are all chasing AI-powered search. The market exists because ChatGPT owns the interface layer but can't prioritize search, and Google has an ad business to protect. If you're building tools that need better search or discovery features, watch this space for API access. (TechCrunch)
Anthropic signed a $1.25B/month cloud services deal with SpaceX to access compute capacity at COLOSSUS and COLOSSUS II through May 2029. That's $15B annually for training infrastructure. Either party can terminate with 90 days' notice. This is the scale of compute required to train frontier models like Claude, and it confirms that training budgets are measured in tens of billions now. (Simon Willison)
Test token speeds before committing to a model. Mike Veerman built a simple HTML app (link) that simulates LLM output speeds from 5 to 800 tokens/second. If you're evaluating models and see "30 tokens/second" advertised, use this to visualize what that latency actually feels like for your users before you build on it.
Automate repetitive desktop tasks with IrisGo. Show the tool how to do something once (ordering supplies, generating reports, filling forms), and it remembers the process for future use. No coding required. If you spend more than an hour a week on the same manual workflow, this pays for itself immediately.
The agent race is heating up. Google shipped Gemini Spark (cloud-based, 24/7, deep Google integration), IrisGo launched desktop automation that learns by watching, and Nvidia's Vera CPU already hit $20B in sales targeting agentic workloads. The pattern across today's stories: agents are moving from demos to infrastructure bets, with billions in compute contracts (Anthropic's $15B/year deal) and hardware designed specifically for agent tasks. Watch for security incidents as sensitive business data flows through these systems at scale. The first major agent breach will reshape how we think about deployment.