Meta, OpenAI and Instinct push personal AI agents from chat into everyday tasks

Meta, OpenAI and Instinct are turning the personal-assistant pitch into shipping products: agents that can browse, connect to apps and keep working after a user closes the chat. The new wave moves the AI contest from answering questions to executing errands and work, with privacy, permission and reliability becoming the central tests. Meta introduced Muse on September 8 as a personal AI agent that can act across daily apps, run in a dedicated Muse Secure VM with its own browser, and ask for approval before sensitive actions such as sending email or making purchases Meta. The company later expanded the product for small businesses with connectors including Asana, Box, Canva, Dropbox, Figma, Intuit QuickBooks, Notion, Shopify, Slack, Stripe and Zoom, while saying nothing publishes, sends or spends without user approval Meta. OpenAI followed on September 29 with dots, always-on agents in ChatGPT powered by GPT-6 Astra, each with its own cloud computer and access to more than 4,000 apps through plugins OpenAI. Dots are rolling out to eligible Pro and Business Premium users, with Enterprise access available as an admin-enabled beta; Pro access excludes the European Economic Area, Switzerland and the UK at launch, and users must create a dot from the desktop app or desktop web OpenAI Help Center. Instinct, meanwhile, describes a text-or-call assistant that connects to applications and devices, including email, messaging, screen, audio and location, and is trained to use a phone and computer on a user’s behalf Instinct. The startup also raised $1 billion at a $10 billion valuation, after a $2.5 billion valuation about a month earlier, underscoring investor expectations that personal agents could become a new consumer interface SiliconANGLE. The competition is important because the category is trying to replace app-by-app labor with persistent digital proxies. OpenAI says dots can work in ChatGPT, Slack and Teams, continue across conversations, and bring results back for review; Meta says Muse can keep working after the app is closed, use Link by Stripe for purchases, and remember user preferences OpenAI Meta. Business users are already a target: Muse for Small Business is positioned around campaign planning, email triage, financial checks and growth plans, while OpenAI says specialist dots are being tested for procurement, invoice processing, email marketing, customer support and commercial contracting Meta OpenAI. The technical shape is converging around isolated computers, connected tools, background work and approval systems. OpenAI says proactive research tasks use read-only tools, cannot directly send messages or change connected apps, and must pass action rules before follow-up steps; it also says dots can still make mistakes OpenAI. Instinct’s privacy policy says the assistant may access sensitive information depending on permissions, including private communications, payment information and health-related material, and warns that autonomous agents can cause unintended payments or communications Instinct. Unofficial public Geekbench runs have also drawn attention to the cloud-computer layer: one hardware analysis said post-launch dots runs ranged from 1,512 to 1,614 single-core and 8,135 to 8,991 multi-core, compared with reported Muse median scores of about 1,041 single-core and 1,394 multi-core Tom's Hardware. Early use cases show both the attraction and the risk. One Instinct test described the assistant monitoring sold-out museum tickets for three days, buying them when availability opened, and rearranging an itinerary around the booking; other reported incidents included a flight cancellation that cost more than $200, an unwanted restaurant reservation with a $200 cancellation fee, and a reservation account temporarily locked after automated requests The Atlantic. Muse has also faced privacy scrutiny after a test account alleged that the agent surfaced context from messages it had not been explicitly allowed to read, including a local message database sync reaching row 187,462 Tom's Hardware. The next phase of the agent race will likely be judged less by novelty than by control: whether these systems can reliably complete useful work, expose what they are doing, respect permissions, and avoid turning convenience into accidental commitments.