Claude Duet Explained: What It Can Do With Your Accounts (and What It Never Will)

What can Claude Duet actually do with your email, your accounts and your card, and what will it never do, even if you say yes? Claude Duet is Anthropic's new proactive agent (research preview, 21 Oct 2026). I don't have access yet, so I read every promise on Anthropic's page and checked what it doesn't say, against what Meta (Muse) and OpenAI (Dots) published for their own agents. What we found (official pages read 21 Oct 2026) • Before Duet sends a message, makes a payment, signs a document or deletes anything, it shows you exactly what it will do and waits for your approval. (Anthropic) • "A few things are off-limits even with your approval, like trading securities or moving money between accounts." Those are examples, not the full list. (Anthropic) • The detailed safety design isn't published yet: Anthropic says it "will share more details … in the coming days". On launch day, Meta published a 20-minute engineering post on Muse's safety, and OpenAI's Dots FAQ lists what your approval can't override. • Anthropic's safety claims (credential handling, sensitive-data classifier, auto mode classifier, penetration tests, public bug bounty) are claims. We have not verified them. • Prompt injection: no company promises zero. Meta: "an open problem". OpenAI: protections "do not eliminate it". Anthropic gives no number. • Two early testers, opposite verdicts on proactivity: Peter Yang ("the most proactive", a check-in every 3 hours) and How I AI ("never pushed us again", slow). Quoted with attribution; we show none of their screens. • Nothing in this video is a recording of Duet. Panels marked "illustration" are our own drawings based on Anthropic's examples. Chapters 0:00 What Claude Duet can do (and never will) 0:34 What Duet is 1:21 The six pieces 2:47 What testers saw 4:38 The safety audit 5:31 The line most videos skipped 7:17 Prompt injection, explained 8:57 The agent wave: OpenClaw to Duet 9:34 Duet vs Dots vs Muse vs Grok Bot (sourced table) 11:50 What it costs you 12:51 How to get in 13:26 Verdict 14:14 The five tests I'll run Sources Anthropic, "Introducing Claude Duet" (21 Oct 2026): https://claude.com/resources/articles/claude-duet Claude Duet product page, FAQ and waitlist: https://claude.com/product/claude-duet Meta, "Introducing Muse" (8 Sep 2026): https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/ Meta, "How We Built Safety Into Muse" (8 Sep 2026): https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse OpenAI, "Introducing dots" (29 Sep 2026): https://openai.com/index/introducing-dots/ OpenAI, Dots privacy, security, and safety FAQs: https://help.openai.com/en/articles/20001529 xAI, "Introducing Grok Bot" (11 Aug 2026): https://x.ai/news/introducing-grok-bot Microsoft, new Copilot with Home, Code and Autopilot (25 Sep 2026): https://news.microsoft.com/source/emea/2026/09/new-microsoft-copilot-brings-home-code-and-autopilot-together/ Google Cloud, "Gemini agent" (8 Oct 2026): https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/gemini-at-work/ OpenClaw repository: https://github.com/openclaw/openclaw Hermes Agent (Nous Research): https://hermes-agent.nousresearch.com/ Simon Willison, "The lethal trifecta for AI agents" (June 2025): https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/ Peter Yang, "Claude Duet Is Here!": https://www.youtube.com/watch?v=D-pz8L1orVI How I AI, "I tested Claude Duet": https://www.youtube.com/watch?v=HYb5w4PfPZM Muse prices ($20 / $100) as reported by Axios and TechCrunch (Meta's post only says "subscription plans"): https://www.axios.com/2026/09/08/meta-debuts-muse-personal-ai-agent Not affiliated with Anthropic, Meta, OpenAI, xAI, Microsoft, Google, Nous Research or OpenClaw. Logos and pageaa captures are shown unaltered to identify the products and sources. Disclosure: researched and written with help from Claude, a model made by Anthropic. Synthetic narrator voice. Original score. ⚡ Wiredeep: the tech you use every day, explained. 👉 Subscribe for the Duet test as soon as I get access. #claudeduet #anthropic #aiagents