Quick answer
An always-on agent holds an ongoing responsibility and keeps working between your messages. It runs on its own compute, watches for events, acts within the permissions you grant, and reports back. OpenAI's Dots (launched September 29, 2026, first one included with Pro and Business Premium), Meta Muse (September 8, now in the US and Canada), and xAI's Grok Bot (early beta since August) are the mainstream examples; Claude's merged chat-and-Cowork experience and ChatGPT's Team Tasks are the same idea in a different shape. The shift from chat to standing responsibility changes what you should ask of a product.
A chatbot waits for you. An always-on agent does not. That one difference explains why every major lab shipped one within a few weeks of each other, and why the interesting questions about them are about permissions and logs rather than intelligence.
How it differs from a chatbot
- Trigger: a chatbot responds to a message; an agent responds to events — a new email, a calendar change, a schedule, a ticket
- Duration: a chat ends; an agent's job continues until you end it
- Compute: Dots each get a dedicated cloud computer; Muse runs tasks in a secure virtual machine
- Access: agents connect to your apps and accounts — Dots to 4,000+ via plugins, Grok Bot by signing in to your tools and operating them as a person would
- Output: not an answer but completed work, plus a report of what it did
The three shipping products
- OpenAI Dots: GPT-6 Astra, plugin ecosystem, reachable from ChatGPT, Slack, and Teams; first Dot included on eligible plans, more to be priced later; EU and UK Pro users waiting
- Meta Muse: a personal agent that runs errands and fills forms in a secure VM; free up to 100 million tokens a week with $20 and $100 tiers; US and Canada
- Grok Bot: an AI teammate that drives your existing tools through their interfaces; early beta, no pricing
Questions to ask before you delegate
- What can it access, and can you scope it to read-only or to specific accounts?
- Does it ask before acting, and can you choose which actions need confirmation?
- Where is the log of what it did, and can you audit it later?
- How is your data isolated from other users' agents? OpenAI's research agents exposed 53 users' images in September; isolation bugs are the risk category for agents
- What does it cost to leave running, and what happens when it hits a plan limit?
Where this goes
Agent payments (Mastercard's Agent Pay), shopping features (ChatGPT Try On), and voice that can act (ChatGPT Voice plugins, Google's Call for Me) all point the same way: the assistant becomes a delegate. The controls around that delegation, permissions, confirmations, logs, and spending limits, are the product features that will separate the agents people trust from the ones they switch off.
Bottom line
An always-on agent is a standing employee with your logins. Hire it the way you would hire one: give it a clear job, limited access, and a habit of reporting back, and expand from there.
