In a Personal Agent, the Person Is Human and the Client Is AI

On September 8, Meta launched Muse in the US, an agent that can send email, book trips, and check out on the web for a user, and keeps working after the app is closed. The basic tier is free, and heavier users can choose tiers at $20 or $100 a month. Three weeks later, on September 29, OpenAI announced dots at DevDay. A dot lives in ChatGPT, runs on GPT-6 Astra, and has its own cloud computer and browser. You can find it in ChatGPT, Slack, and Teams, and Pro and Business Premium users each get one. In describing internal use, OpenAI said a dot notices on its own an invoice that was missed, prepares it, and hands it to the user to approve before sending. Add Grok Bot, which went into beta in August, and in a little over a month several large companies each have a personal agent, with launches focused on which tasks it can handle for the user. On October 1, Decagon, which makes AI for customer service, published a blog post addressed to businesses. Decagon argues that agents like Muse and dots will soon come to companies on users' behalf to place orders, cancel subscriptions, and negotiate prices, and companies must plan how to receive them. It proposes a protocol called PACT (Personal Agent Consent & Trust), built on OAuth 2.0 and the Agent2Agent protocol, in which a visiting agent proves to the company which person it acts for and how far that person allows it to go. In Decagon's picture, the company's systems deal directly with the agent and check with it a person's identity and authorization. I think this is closer to how personal agents will be used than what the launches describe: "personal" refers to a person, but most of what uses the information about that person will be AI. Booking a plane ticket Booking a plane ticket means opening the airline site, comparing prices, choosing a seat, and paying. Those steps are the same whoever books, and a different model or a different company's agent can do them. What is specific to the traveler is something else: whether they prefer the aisle or the window, the budget for this trip, whether they can accept a connection, and whether to buy the ticket now, a last question the agent will probably come back to ask. For chores like booking and filling in forms, the gaps between agents won't last long. In the piece on cloud computers I said that capability is the easiest line to catch up on. These chores will also be split among more agents. When OpenAI introduced dots it said there will be several dots working together for a user, and companies will have dedicated dots, each with its own identity, credentials, and tools.