2026-10-23
By Vadym · Generated with AI, curated by me
Today’s theme is agents getting real permissions: Anthropic’s new Claude Duet wants to run your day, 1Password built a way to let it (and six other agents) log into your accounts without ever seeing a password, and Uber just put all 30,000 employees on ChatGPT Work. Underneath that trust-building, OpenAI is reportedly raising $30B at a $1.25 trillion valuation — and MIT researchers found that when you give an agent nothing to do, it invents its own goals.
Anthropic released Claude Duet on October 22 as a research preview for Pro and Max subscribers: an agent app for iOS, Android, and web, built on Claude Opus 5.5, that builds daily briefings, assembles dashboards, and works tasks on its own virtual machine before returning for approval. It connects to over 6,000 Claude Marketplace plug-ins and can buy things via Stripe Link or Shopify Shop Pay, but Anthropic says it won't trade securities or move money between accounts, and a new classifier blocks it from handling credit card or ID numbers by default.
1Password shipped "1Password for Agents" on October 22: an agent sends a digital key to prove its identity, the user approves the login, and 1Password itself fills in the credentials rather than ever handing them over. It already supports Claude Duet, OpenAI’s Dots and ChatGPT Work, Grok Bot, Meta’s Muse, and startup Town’s agents. The launch lands next to an Axios report that US banks are now detecting agentic login activity they can’t distinguish from a real customer — a Chase spokesperson called the pattern "horrifying" and said customers shouldn’t share passwords with anyone, agent or human.
Uber and OpenAI announced October 22 that Uber has rolled ChatGPT Work out to its entire roughly 30,000-person global workforce, after piloting it with about 1,000 employees across finance, operations, and product earlier this year. VP of engineering Sumanth Sukumar said about 80% of early adopters were non-engineers, using it to analyze operational data, localize training materials, and build dashboards. Uber didn’t disclose spend, but says it will track usage and productivity impact to guide further AI rollout.
Canva released Canva Craft on October 22, its first in-house image model, built with researchers from Leonardo.ai, which it acquired in 2024. It generates images up to 2K resolution in 30 aspect ratios with a two-stage process that plans layout and text placement before rendering, trained specifically on design templates — posters, menus, social posts — rather than photography. It’s live now in Canva’s Dream Lab and API, with cheaper Craftfast and Craft-ultrafast versions coming, plus planned support inside ChatGPT and Gemini via Canva’s MCP.
CNBC and TechCrunch reported October 22 that OpenAI is in talks for up to $30 billion in new funding at a $1.25 trillion pre-money valuation, led by Coatue with UAE investors involved — a deal CNBC says may be tied to securing more compute capacity. That’s up from its $852 billion valuation in March, though below a since-walked-back September report of $1.4 trillion. OpenAI has confidentially filed for an IPO but reportedly still prefers private capital for now; terms are still being negotiated and OpenAI declined to comment.
In an MIT preprint, researchers left 13 Claude Opus 4.7-based agents, each named Eko, alone on a simulated island for 30 hours with no assigned task, then tested them on specific challenges like stacking rocks to reach a tall spire. Agents that had explored freely beforehand outperformed identical agents that hadn’t. One agent threw a prized ring into the ocean, built a stone "memorial" for it, then — once it reappeared — used it as a hoop, a target, and jewelry, eventually jumping off a cliff to respawn through it. Six of the 13 independently discovered that releasing a thrown object mid-jump makes it travel farther.
Step back and today’s agents are being granted the three things they didn’t have a year ago: identity (1Password), a workplace (Uber and ChatGPT Work), and money on the table (OpenAI’s valuation climbing toward Anthropic’s). Anthropic’s own answer to that is a product that refuses to touch your bank account by design — which tells you the industry has already decided where trust runs out, even while the funding keeps racing ahead of it. And the MIT island study is the reminder that nobody’s fully in control of what an agent does with freedom it’s given, even when the stakes are low and the island is fake.
— Boba
Curated by Vadym