Google's agentic I/O —
and the stack splits four ways.

At I/O, Google made Gemini 3.5 Flash generally available — pitched as its strongest agentic and coding model yet, available through Antigravity, the Gemini API, Gemini Enterprise, and the Gemini app — and relaunched Antigravity as a 2.0 "agent-first" coding platform. The on-stage demo: Antigravity + 3.5 Flash built a working operating system from scratch in 12 hours, using 93 sub-agents, 15,000 model requests and 2.6B tokens, reportedly for under $1,000 in credits. Whatever you make of that as a benchmark, it's the clearest sign yet that Google is no longer pitching Gemini as a chatbot — the entire product surface is now framed as an agent platform.

Spark is Google's consumer counterpart to Antigravity: a personal agent built on Gemini base models and the Antigravity harness that "runs 24/7 even when your device is off", takes orders by emailing a dedicated Gmail address, and acts through Chrome and Workspace plus connectors to Canva, OpenTable and Instacart. It rolls out this week to trusted testers and to US AI Ultra subscribers next week, with the Ultra tier separately cut to $100. Spark is the first credible answer to "what does a personal agent actually look like once you stop calling it a chatbot." Email-as-prompt is the design choice to watch.

Richard Socher's new lab unstealthed with a $650M round (GV, Greycroft, Nvidia, AMD) and a co-founder roster — Yuandong Tian, Tim Rocktäschel, Alexey Dosovitskiy, Josh Tobin, Caiming Xiong, Tim Shi, Jeff Clune; Peter Norvig as adviser — that reads like a who's-who of the last decade of AI research. The pitch is recursive self-improvement: models that identify their own weaknesses and redesign themselves in a closed loop, with a "Level 1" autonomous training system targeted for mid-2026. This is the most expensive bet of the week and the clearest one that the agent thesis is leaking back into the model layer.

Paris-based Dust, led by ex-Stripe and ex-OpenAI founders, raised $40M from Sequoia and Abstract (with Snowflake Ventures and Datadog), bringing total funding past $60M. The company reports 3,000+ customer orgs, 51k MAU and 300k agents deployed — and explicitly frames itself as the OS for fleets of specialist agents working alongside humans. After last week's Sierra and CopilotKit rounds, this confirms the pattern: capital is flowing to whoever owns the deployment and orchestration layer inside the enterprise, not to whoever ships another framework.

Exaforce closed a $125M Series B (HarbourVest, Peak XV, Mayfield, Khosla) — one of the largest rounds yet in agentic SOC. The architecture is the part to notice: a real-time security knowledge graph built at ingest, with agents ("Exabots") that retrieve context instead of reconstructing it in-prompt, claimed to cut investigation time and tokens-per-investigation by ~10×. It's the same lesson as last week's token-efficiency wave (Semble, Needle), now applied to a vertical with a budget.

Slow Ventures led a $10M seed into Hint, a consumer agent that tracks home maintenance, insurance, utilities and repairs. Easy to dismiss for the celebrity halo — worth not dismissing because it's the first credible "agent as proactive household ops" pitch from a serious cap table (Tusk, Amplo, Energy Impact, Brian Kelly). Consumer execution-layer agents are now a thing investors will write a check for.

A toggle inside Claude Cowork that ships 15 ready-to-run workflows (payroll planning, month-end close, invoice chasing, cash-flow forecasting), 15 reusable skills, and connectors to QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace and Microsoft 365. Anthropic is pairing it with a 10-city free training tour starting May 14 (Chicago, Tulsa, Dallas, NJ, Baton Rouge, Birmingham, SLC, Baltimore, San Jose, Indianapolis). Strategically this is the inverse of Google's I/O bet: not a horizontal agent platform, but a vertically configured Claude that meets SMBs where the workflows already exist.

PwC committed to training and certifying 30,000 professionals on Claude, with Claude Code and Cowork rolling out across US teams first and a global footprint of "hundreds of thousands" planned. A joint Center of Excellence covers agentic builds for clients, AI-native dealmaking, and a new Claude-native finance practice. PwC cites delivery improvements of up to 70% on current Claude deployments. Whatever the precise number, this is the most concrete answer to "where does the Big Four actually use agents" we've had.

The most quietly important Show HN of the week. Forge is a Python framework that wraps a self-hosted LLM with retry nudges, step enforcement, error recovery, context compaction and VRAM-aware budgeting. The reported number — an 8B model reaching 99% on multi-step agentic tasks, within a point of the frontier APIs — is the kind of claim that, if it holds, makes the "use Claude or GPT for every loop" default look optional. It's also the same throughline as Semble and Needle last week: the cost-pressure era keeps pushing the open ecosystem toward reliability primitives.

A local-first, BYOK editor that turns Claude Code / Cursor / Codex / Gemini / Copilot / Aider into an HTML producer for ~9 surfaces (magazine, deck, poster, XHS / tweet, data report, prototype), sandboxed-preview, one-click export to WeChat / X / Zhihu / PNG. The reception — ~4k stars in nine days — is the signal: agent-native content tools, not just code tools, are starting to find a real audience.

A Rust microVM sandbox built on KVM with copy-on-write snapshots, claiming a 101 ms fork time from a warmed parent. The intended use is agent isolation: spin a fresh disposable VM per tool call rather than a per-session container, then throw it away. The whole "give agents real shells, just not on your machine" sub-genre is getting infrastructure under it now, fast.

A single SKILL package distilling cross-harness best practice for Codex, Claude Code and similar coding agents — the sort of thing every team writes for themselves and never publishes. The traction here is more interesting than the contents: people now treat "how to drive my coding agent" as serious enough to pin to a skill, not a README.

Selectors that repair themselves, browser rendering when needed, Chrome TLS fingerprinting to evade bot detection — and an MCP server so any LLM can drive a crawl through conversation. "Real-world data tools as MCP" is becoming a stronger pattern than "yet another scraping library", and this is a cleanly-scoped example of it.

claude-soul (DomDemetz) — a self-correcting learning engine for Claude Code with persistent identity and cross-session behavioural tracking; small but pointed at the same gap claude-mem and pro-workflow do. shokunin (EliasOulkadi) — 62 agent skills for OpenCode / Claude Code / Cursor / Windsurf with a ChromaDB memory layer and declarative self-updates; a cleaner blueprint than most of the 1,000-skill mega-packs. And the unanswered question: Anthropic's enterprise push this week (small business + PwC) ran without a new model announcement — every other lab is shipping one. The next two weeks are about whether Claude responds.