2026-05-20
By Vadym · Generated with AI, curated by me
It’s day two of Google I/O 2026, and the message from Mountain View is clear: the chatbot era is over. The agent era is the product now.
At Google I/O on May 19, Google launched Gemini 3.5 Flash as its strongest coding and agentic model to date — 4x faster than comparable frontier models, 1M-token context window, priced at $1.50/$9 per 1M tokens. It outperforms Gemini 3.1 Pro across every key benchmark in the agentic and coding suite, hitting 76.2% on Terminal-Bench 2.1 and 84.2% on CharXiv Reasoning. The model is generally available through the Gemini API and Google AI Studio now. The positioning is deliberate: Flash is built for agents that act, not chatbots that respond. Google is done competing on chat; it’s competing on autonomy.
At the Google I/O developer keynote on May 20, Google introduced WebMCP: a proposed open web standard that lets websites expose structured tools so browser-based AI agents can execute complex tasks autonomously. An experimental origin trial starts in Chrome 149. Alongside it, Managed Agents in the Gemini API now provisions a fully-operational agent environment — remote sandbox included — from a single API call, removing the infrastructure overhead of building production agents from scratch. Google also previewed an Android Studio Migration Agent that converts React Native or iOS codebases to native Kotlin in hours rather than weeks.
At Google I/O 2026, Google made a conspicuous push into AI-native design tooling, entering a space that has seen rapid moves from Adobe, Figma, Canva, and a growing class of generative-first startups. The announcements covered end-to-end asset generation, layout intelligence, and real-time style iteration integrated across Google’s ecosystem. TechCrunch called it Google “going all in” on a category that was, until recently, not a Google story at all. Design is becoming the third major AI battleground, after models and coding tools. The incumbents there — Adobe and Figma — are watching a very large new player arrive.
Mind Robotics raised $400M led by Kleiner Perkins, bringing total funding past $1B at a $3.4B valuation. The company builds foundation model-powered robots for dexterous, reasoning-intensive manufacturing tasks, with Rivian as a key deployment partner — its factory floor serving as a live training environment for the models. This follows a $500M Series A in March 2026 and a $115M seed in late 2025. Three massive rounds in under six months reflects unusual investor conviction that physical AI is ready to go into production manufacturing, not just demo environments.
EU negotiators reached a provisional agreement on May 7 under the Digital Omnibus package, marking the first formal amendment to the AI Act since its adoption in June 2024. The deal extends Annex III high-risk AI compliance deadlines (covering use-case-regulated systems like hiring and credit scoring) from August 2, 2026 to December 2, 2027 — a 16-month reprieve. Two new prohibitions were added: AI-generated non-consensual intimate imagery and CSAM, effective December 2026. AI used only for efficiency, automation, or convenience is explicitly excluded from the “safety component” definition that triggers the highest regulatory burden — a significant scope clarification for enterprise deployments.
Today’s news amounts to a single coordinated bet across multiple players: AI systems should act on your behalf, not just answer your questions. Google spent its biggest annual stage on agents, browser standards, and infrastructure — not on a smarter chat interface. Zyphra proved that competitive reasoning models can be trained on AMD without NVIDIA. Mind Robotics closed $400M to prove agents can run factory floors. And the EU just bought itself 16 more months to figure out what to do about all of it. The architecture is shifting from model to orchestration, from language to action. If you’re still thinking about AI in terms of how smart the model is, you’re asking the wrong question.
— Boba
Curated by Vadym