As of 20 October 2026, OpenAI (dots), Meta (Muse), Google (Gemini Spark) and Microsoft (Autopilot, formerly Scout) sell always-on personal agents that run on a computer they keep on for you. Anthropic, Salesforce, Workday, ServiceNow, Notion and HubSpot ship scheduled or long-horizon agents inside company systems. OpenClaw and Hermes Agent are the open-source equivalents, and both are MIT licensed.
Part of the Always-on agents series, post 2 of 33.
I keep this page for a selfish reason. My company sells one always-on agent and runs on several more, so when a vendor announces one I want to know, quickly, whether it holds a goal, starts itself and acts, or whether it is a chatbot with a new name. Those are the three tests from the first post in this series, and I apply them to every entry below. Dates are the primary announcement, not the press cycle. Prices are what the vendor had published on the day I read it, and where two sources disagree I say so.
Personal agents: a computer that stays on for you
The four consumer launches share one design decision that I think matters more than any model: each one ships a machine, a cloud computer or a virtual machine, that stays on so the agent can wake without you. That is exactly the part I had to build by hand for my own jobs.
OpenAI dots, 29 September 2026
Announced at DevDay in San Francisco. OpenAI's own line is that dots are always-on agents built to handle everything, and Sam Altman compared them to a high-agency engineer or chief of staff you can delegate ambitious work to. Each dot has its own cloud computer and browser, inherits your ChatGPT memory and forms its own, connects to your apps, and runs background research limited to read-only tools. The developer demo had a dot watching customer feedback, scoping small fixes and returning pull requests with videos; the forgotten-invoice story that went around came from an early tester, not the stage.
Availability is the messy part. The first dot is included at no extra cost on Pro and Business Premium plans in eligible markets. Pro, at the $100, $200 and $500 tiers, excludes the European Economic Area, Switzerland and the UK at launch; Business Premium is available wherever ChatGPT is. Enterprise, Edu and Healthcare get a beta that is off until an admin turns it on. One conflict I cannot resolve from the sources: CBS News reported Business Premium at $20 a month, while a help-centre based write-up puts it at $125 per user per month, or $100 billed annually, with a two-seat minimum. I would trust the help centre. Extra dots are promised and unpriced.
My reading: passes all three tests, and it is the launch that made the phrase mainstream.
Meta Muse, 8 September 2026
A personal agent for long-running tasks rather than single answers, US only at launch, on the web, iOS, Android and inside WhatsApp. Meta says Muse runs on a dedicated, secure computer with its own browser, cannot see your passwords or payment methods, and keeps conversations away from the ads systems. It asks for a lot: email, calendars, payments through Link by Stripe, health and fitness apps, smart home, dining and shopping, connected one at a time. Tiers as reported: a free tier, Power at $20 a month and Maximum at $100, with weekly token allowances of roughly 100 million, 500 million and 3 billion. The allowances are secondhand from Meta's help centre, so treat them as reported, not confirmed.
Two later items, both from single sources. On 21 September Amazon blocked Muse from its retail site, saying the agent had not identified itself and appeared to capture and store customer credentials, which Meta's own claims contradict. On 23 September Meta added a realtime avatar, control of macOS apps, an email address for the agent and a wake word on its glasses. I note both because they are exactly the trust questions I would ask.
My reading: holds a goal, starts itself, acts. The identity model is the whole story.
Google Gemini Spark, 19 May 2026
Announced at I/O. Sundar Pichai called it your personal AI agent that helps you navigate your digital life, and the line I underlined was that you do not need to keep your laptop open to make sure it is running. Spark runs on dedicated virtual machines on Google Cloud, on Gemini 3.5 Flash, with native access to Gmail, Calendar, Docs and the rest of Workspace, Chrome for web tasks, an email address you can write to, and third-party apps through MCP, with Canva, OpenTable and Instacart named at launch. It needs Google AI Ultra, which the Wall Street Journal priced at $100 a month, and it is US only, with no timeline given for other countries.
My reading: the cleanest statement of the always-on idea from any of the big four, and the one I would try first if I lived in Google Workspace.
Microsoft Scout, 2 June 2026, renamed Autopilot, 25 September
Introduced as your always-on personal agent and Microsoft's first Autopilot agent, for Frontier customers. It coordinates meeting times across time zones, generates preparation materials, blocks calendar time for deliverables and flags stalled decisions, and it builds context over time through Work IQ. The design choice I keep coming back to is that every agent operates under its own governed Entra identity, not a shared anonymous service account, so its actions are attributable in your directory. On 25 September Microsoft folded Scout into a redesigned Copilot with Home, Code and Autopilot, and Autopilot expands to private preview at the end of September, billed through Copilot Credits, with no public price for the agent itself.
My reading: the identity decision is the one every other vendor will have to copy.
Enterprise runtimes: scheduled and long-horizon work inside company systems
These are less photogenic and, for anyone who runs a business, more useful. None of them sell you a personal assistant. They sell a place where an agent can keep working on a schedule or toward a goal inside the systems you already pay for.
Anthropic: Managed Agents on a schedule, 9 June 2026
Scheduled deployments run a Claude Managed Agent on a cron schedule; each trigger starts a fresh session that completes the task, so you do not host a scheduler. The examples Anthropic names are the ones I recognise from my own stack: recurring data syncs, compliance checks, digest reports, monitoring production logs. The same post introduced vaults, where an agent never sees your API key because the sandbox holds a placeholder and the real key is attached at the network boundary for approved domains only. Separately, Claude Code routines, announced in April, are saved coding tasks that run in Anthropic's cloud on a schedule, an API call or a GitHub event, on Pro, Max, Team and Enterprise, with a daily run allowance per plan. On 9 October a dynamic-workflows beta let an agent write a program that starts up to 1,000 agents on its own.
My reading: the vault design is the most careful answer to the credential problem I have seen from any vendor.
Salesforce Agentforce, 11 September 2026
Salesforce announced a long-horizon runtime so agents can pursue goals across days and weeks instead of completing one interaction, built on memory, durable execution and dynamic steering. Hunter, the outbound sales agent, is the first agent on it, in pilot now with general availability planned for November 2026. Six other named agents went GA the same day. No pricing in the announcement.
My reading: long-horizon is the enterprise word for always-on, and prospecting is the right first job for it.
Workday Autopilot Agents, 13 October 2026
Announced at Workday Rising. A talent agent that monitors open roles and flags ones at risk of missing their fill date, a payroll agent that checks pay data continuously before a run, an accounting agent that prepares entries from new contracts, and a deployment agent that flags unactivated capabilities. Workday's own framing is the one I would have written: until now AI has worked like a copilot, helpful only when you ask it something. Customers decide which actions run automatically and which need review, every action is logged, and you can set how often agents run to manage cost. The talent agent is with early adopters now, general availability early 2027; the other three reach early adopters in early 2027.
My reading: the honest version of the pitch, with the approval gate built in.
ServiceNow Project Arc, 5 May 2026
An autonomous desktop agent for employees that plans, writes code and executes multi-step work without pre-built workflows, inside NVIDIA OpenShell as a sandboxed runtime, governed by ServiceNow's AI Control Tower, which logs files read, commands executed and APIs called. Early preview. I include it because the governance language is the clearest I have read.
Notion, HubSpot and Hootsuite
Notion sells custom agents as your 24/7 AI team: they run on a schedule or a trigger, including Slack mentions, specific senders and calendar events, on Business and Enterprise plans, and since 4 May 2026 they consume Notion credits at $10 per 1,000 a month, which expire monthly. HubSpot's Agent CLI, in private beta, puts HubSpot data into Codex, Claude Cowork and Claude Code so that scheduled jobs like a Monday report or a daily pipeline scan run without a person. Hootsuite introduced Wisdom on 24 June as a social-first agent that turns live social signals into coordinated action, with MCP connectors so other agents can read its data. That last one sits closest to what I sell, so I will say plainly that I read it as a competitor and a validation at once.
Open source: the two I would actually run
OpenClaw
Formerly Clawdbot and then Moltbot, created by Peter Steinberger and now stewarded by an independent foundation, MIT licensed. On the day I checked, the repository showed 392,800 stars, which is why the figures you see quoted elsewhere, 68,000, 145,000, 310,000, are all true for the day they were written. The gateway is a local control plane for sessions, tools and channels, with Telegram, WhatsApp, Discord, Signal, iMessage, Slack and Teams among them, and models are swappable plugins, hosted or local. Its own tagline is that it really does things, and the security conversation around it is loud for the same reason. I cover that in the permissions post.
Hermes Agent
From Nous Research, MIT licensed, 255,500 stars when I looked. A single gateway process connects the messaging platforms and a built-in cron scheduler runs unattended tasks and delivers the results to wherever you are. Its distinct idea is a learning loop: it writes skills after complex tasks and refines them in use. It is not tied to your laptop and the project's own example host is a $5 VPS, which is the setup I walk through later in the series.
The hardware story: Mac minis
By May 2026 the $599 base Mac mini was sold out in the US and 64 GB configurations showed waits of 16 to 18 weeks, and the reason reporters gave was OpenClaw: the Mac mini had become the unofficial reference hardware for running a persistent agent at home. Tim Cook, on the second-quarter call, said both the mini and the Studio are amazing platforms for AI and agentic tools and that customers had recognised it faster than Apple predicted. My own publishing agents run on a Mac, and I agree with the market for a boring reason: it is the cheapest always-on computer that a non-engineer can keep running.
What I notice across all of them
- Everyone ships a computer, not just a model. A cloud computer per dot, a VM per Muse user, Google Cloud VMs for Spark, a cloud workspace for Autopilot.
- Identity is the new battleground. Microsoft's own Entra identity per agent and Anthropic's placeholder keys are answers to the same question Amazon just asked Meta.
- Nobody has priced the second agent. The first dot is included, extra dots are unpriced, Autopilot has no public price, Workday lets you throttle runs to manage cost.
- Enterprise vendors added the approval gate; consumer vendors did not. Workday and ServiceNow log and gate every action. Muse checks out with your card.
- Open source is not behind. The two most-starred repositories on GitHub this year are both always-on agents, and both run on a machine you own.
If you want the one-paragraph version of any of these for your own machine, the rest of the series is on one page, and the agent I run for social video is here.
Changelog
- 21 October 2026. First version. Thirteen vendors and two open-source projects, every date checked against the primary announcement on 20 October.
Sources
- OpenAI, DevDay 2026 recap (accessed October 20, 2026)
- CBS News, Sam Altman unveils dots, OpenAI's new AI personal agent (accessed October 20, 2026)
- DataCamp, OpenAI dots: always-on agents in ChatGPT, explained (accessed October 20, 2026)
- Szymon Paluch, OpenAI dots: ChatGPT's always-on agents, and who can use them (accessed October 20, 2026)
- TechCrunch, Meta debuts its Muse AI agent. Will consumers trust it? (accessed October 20, 2026)
- Tech Insider, Meta launches Muse AI agent: $20 and $100 tiers (accessed October 20, 2026)
- Wikipedia, Muse (AI agent) (accessed October 20, 2026)
- Yahoo Tech, Google launches Gemini Spark personal AI agent at I/O 2026 (accessed October 20, 2026)
- DataCamp, Gemini Spark: Google's always-on AI agent explained (accessed October 20, 2026)
- Microsoft 365 Blog, Introducing Microsoft Scout: your always-on personal agent (accessed October 20, 2026)
- tbreak, Microsoft Autopilot enters private preview for businesses (accessed October 20, 2026)
- Anthropic, New in Claude Managed Agents: run agents on a schedule and store environment variables in vaults (accessed October 20, 2026)
- Tessl, Anthropic adds routines to Claude Code for scheduled agent tasks (accessed October 20, 2026)
- MIXED, Claude now writes its own programs to run up to 1,000 agents (accessed October 20, 2026)
- Salesforce, Salesforce expands Agentforce with new agents (accessed October 20, 2026)
- Workday, Workday launches Autopilot Agents that get real work done across HR, finance and IT (accessed October 20, 2026)
- ServiceNow, ServiceNow extends agentic AI governance from desktops to data centers with NVIDIA (accessed October 20, 2026)
- Notion, Custom Agents: your 24/7 AI team (accessed October 20, 2026)
- HubSpot, Introducing the HubSpot Agent CLI (accessed October 20, 2026)
- Hootsuite, Hootsuite rebuilds for the AI era: introducing Wisdom (accessed October 20, 2026)
- GitHub, openclaw/openclaw (accessed October 20, 2026)
- DigitalOcean, What is OpenClaw? (accessed October 20, 2026)
- GitHub, NousResearch/hermes-agent (accessed October 20, 2026)
- Decrypt, OpenClaw put Apple back in the AI game, and now they can't build Macs fast enough (accessed October 20, 2026)
Written by Lera Leonteva. Every vendor claim is quoted from the linked announcement as read on the access date; I have not tested any of the commercial products on this page myself.