Always-On Agents: Dots vs Grok Bot vs Muse

Always-on agents keep working after you close the app. Each one runs on a computer that keeps its files and memory, starts tasks on a schedule or when something happens, and messages you when it needs a decision. We compared five of them: OpenAI Dots, Grok Bot, Meta Muse, Hermes Agent and OpenClaw.

Always-on AI agents comparison

*Nous Research also hosts Hermes Agent on Hermes Cloud. **The software is ücretsiz. You pay for the server and the model usage.

What are always-on AI agents used for?

You give an always-on agent a standing job instead of a one-off question. The examples vendors publish fall into four kinds of work:

• Daily briefings: Hermes Agent’s tutorial builds a bot that researches chosen topics at 8:00 AM and sends a summary to Telegram or Discord.
• Watching for changes: Meta’s example Muse read school emails and the district website, added key dates to a family calendar and flagged a sports tryout 12 hours before registration closed. OpenClaw’s docs show an inbox check every 30 minutes.
• Personal tasks you approve first: a dot finds another time for a meeting that clashes with your child’s recital, or orders dinner on DoorDash once you approve the total.
• Team work: a SpaceXAI sales bot updates the CRM from call transcripts and drafts follow-ups, and an ops bot processes invoices received in Gmail.

Feature comparison

Separate agents each keep their own instructions, memory and routines, so a team can give each one a role. Hermes Agent and OpenClaw also keep separate logins. Grok Bot’s bots share one computer and every login on it.

Hermes Agent, OpenClaw and Muse store memory as text files you can open and edit. Dots keeps memory on OpenAI’s side, and clearing it means deleting the dot. Cursor documents no way to view or edit Grok Bot’s memory.

Always-on AI agent pricing

*Billed annually; $125 billed monthly. All prices are US list prices on September 30, 2026.

• Dots: usage does not count toward plan allowances for the first month after launch. OpenAI has not published the terms after that.
• Grok Bot: Cursor publishes relative usage tiers, not numbers, and warns that an hourly routine or a trigger on a busy Slack channel “can use a week of usage in a day.”¹
• Muse: Power includes 500 million Muse tokens a week and Maximum 3 billion. Meta does not say what a Muse token measures.
• Hermes Cloud: a medium server costs $0,56 a day while running, about $17 for a 30-day month, and $0,03 a day when stopped. Model usage is billed separately.

Top always-on AI agents

OpenAI Dots

OpenAI announced dots at DevDay on September 29, 2026, and is rolling them out gradually. Each dot gets its own Linux computer and Chrome browser in OpenAI’s cloud, connects to over 4.000 apps through ChatGPT plugins, and can buy with a card saved on a merchant site after you approve.²

Each user gets one dot. OpenAI says users can add more later but hasn’t priced them.

Key features:

• Scheduled tasks, plus follow-ups the dot schedules for itself.
• Reads connected apps in the background with read-only tools.
• Delegates coding tasks to Codex.
• Takes voice calls.

Grok Bot

Grok Bot is SpaceXAI’s app of named AI agents, in beta since August 11, 2026. SpaceX acquired Cursor on August 14, and Cursor builds, hosts and bills the product. Each user gets one cloud computer in the US with a browser, terminal and files, shared by all of that user’s personal bots.³

Up to six bots can share a group chat and pass work to each other. Cursor picks the model for each task, and its documentation doesn’t name one.

Key features:

• Team Bots, shared agents with their own Slack handle, in public beta on Teams and Enterprise plans since September 28, 2026.
• Sign-in tokens for connected apps stay on Cursor’s servers, outside the bot’s computer.
• A separate review model checks risky actions before they run.
• Records up to 10 minutes of your screen and drafts a skill from it, a saved set of instructions the bot can reuse.

Meta Muse

Meta launched Muse on September 8, 2026, for adults in the US, and it has since reached Canada. Each user’s Muse runs on a dedicated cloud virtual machine and “keeps working after people close the app.”⁴ You talk to it in the Muse app, on the web or in WhatsApp.

Muse targets shopping and household errands. It pays with a single-use card from Stripe’s Link, and every purchase needs your approval with the exact details. Training on your data is on by default, with an opt-out.

Key features:

• Recurring tasks run until you cancel them.
• A separate agent approves every connector action and outbound request.
• An activity log lists past and planned actions.

Hermes Agent (Bot Mode)

Hermes Agent is an open-source (MIT) agent from Nous Research, public since February 25, 2026, with 250.000+ GitHub stars on September 30.⁵

Bot Mode, on by default in the desktop app since August 31, turns Hermes profiles into named bots, each with its own model, memory, skills and routines.

Key features:

• 28 messaging adapters, including Telegram, Slack, WhatsApp, Signal, email and iMessage bridges.
• Any of 45 listed model providers, or a local model through Ollama or vLLM.
• Writes its own skills after multi-step tasks.

OpenClaw

OpenClaw is an open-source (MIT) personal agent with 390.000+ GitHub stars on September 30, 2026, the most of any agent here. The OpenClaw Foundation, a US non-profit announced on July 8, maintains it. Reuters reported that Meta modeled Muse on OpenClaw.⁶

A heartbeat wakes the agent every 30 minutes by default (hourly with Anthropic sign-in) to check for anything that needs attention. The project sells no hosting. Moonshot AI’s Kimi Claw sells a hosted OpenClaw that stays online.

In our experience, OpenClaw is hard to set up without prior server and configuration-file experience. Its configuration reference spans 12 pages, covering chat channels, agents, tools, the gateway, hooks, and secrets.⁷

Key features:

• 20+ chat apps, including WhatsApp, Telegram, Signal, iMessage and Teams, plus phone calls through Twilio.
• Anthropic, OpenAI, Google, xAI, OpenRouter and many other model providers, or a local model.
• Several agents per installation, each with its own workspace and logins.

We compare OpenClaw with Hermes, NanoBot, ZeroClaw and PicoClaw in OpenClaw alternatives.

Kıyaslamalarımızı ve veri odaklı içgörülerimizi kaçırmayın. Düğme Google'ı açar; AIMultiple'ı seçmeniz, Google arama sonuçlarında AIMultiple'ı daha sık görmek istediğinizi onaylar.

Tercih edilen kaynak olarak ekle

How to choose an always-on AI agent?

Two questions decide most of it: who runs the computer, and who pays for the model.

Hosted agents: Dots, Grok Bot and Muse

The vendor runs the computer, so the agent keeps working when your laptop is closed, and model usage is part of the plan. If you already pay for ChatGPT Pro, a Cursor plan or SuperGrok, the agent comes at no extra charge. We picked Grok Bot for that reason: our SuperGrok plan already included it, and we ask it for a daily update on our projects.

Among the three, Dots ties one agent to ChatGPT, Codex and Slack. Grok Bot runs many role-based bots for a team and bills usage past the weekly allowance at token cost. Muse is a consumer app with a ücretsiz tier that you can text on WhatsApp.

Self-hosted agents: Hermes Agent and OpenClaw

Both run wherever you install them. On a laptop, scheduled jobs pause while it sleeps and run once when it wakes. OpenClaw’s FAQ calls a dedicated host “not required, but recommended for reliability and isolation.”⁸ Hermes’s security guide advises a separate machine or VM “for maximum security.”⁹ OpenClaw’s own setup guide uses a 1 GB DigitalOcean server at about $6 a month, and Hermes Cloud costs about $17.

The model is billed separately and can cost more than the server. OpenClaw’s heartbeat runs a full agent turn every 30 minutes, and each turn can resend about 100.000 tokens of conversation history.¹⁰ That adds up to 4.8 million input tokens a day, about $19 at Anthropic’s list price for Claude Opus 5.5 ($4 per million).

OpenClaw’s isolated-session setting cuts each heartbeat to 2.000 to 5.000 tokens. Claude subscriptions have not covered OpenClaw since April 4, 2026, so Claude usage there is billed as extra usage or through an API key.¹¹

In return, you pick the model, including a local one that runs on your own hardware, and you keep memory in files you control.

Hermes Agent or OpenClaw?

OpenClaw has the larger community: 390.000+ GitHub stars and a public skill registry. Hermes reached 250.000 stars seven months after its public release, and Nous runs a hosted version, Hermes Cloud. The New Stack puts the difference this way: “One prioritizes ecosystem reach, the other deep learning over time.”¹²

Each project imports the other’s setup (hermes claw migrate and openclaw onboard –import-from hermes), so switching later is cheap.

Background task agents

These tools run tasks you schedule or trigger in the cloud, so work continues even when your laptop is closed. Unlike Dots, Grok Bot and Muse, none is documented to keep a persistent computer for each user by default. Several of them operate a browser or desktop the way a person would; we explain how that works in computer use agents.

• Gemini Spark (Google): tasks in the Gemini app on Google Cloud machines, on a schedule or an event such as a flight delay, up to 15 at once. Tasks that use your local Chrome browser need the device awake. From $19,99/mo.
• Microsoft Copilot Cowork: Microsoft 365 tasks started by a schedule, an email or a Teams message. Each task runs in a temporary workspace that is deleted when it finishes. $30/user/mo Copilot license paid yearly, plus usage.
• Claude (Anthropic): cloud Cowork sessions, scheduled tasks and Claude Code routines keep running with the laptop closed, except for steps that need local files or apps. From $20/mo.
• Perplexity Computer: each task, or each Automation triggered by Slack, Gmail or GitHub, gets a fresh cloud machine. Pro $20/mo plus credits.
• Manus: scheduled and triggered Automations on temporary machines; a Cloud Computer that stays on costs $30 or $50/mo extra.

To build an agent of your own instead, see our comparison of AI agent tools.

Bu araştırmayı kaynak gösterin

Yayınlayacağınız yere uygun formatı seçin. Bağlantılı sürümü CMS'inize yapıştırmak, geri bağlantıyı korur.

@misc{kalelioglu2026,
  author = {Kalelioğlu, Berk},
  title  = {{Always-On Agents: Dots vs Grok Bot vs Muse}},
  year   = {2026},
  month  = sep,
  howpublished    = {\url{https://aimultiple.com/always-on-agents}},
  note   = {AIMultiple. Erişim tarihi: 30 Eylül 2026}
}

Referans Linkleri

1.

https://cursor.com/help/grok-bot/routines

2.

https://openai.com/index/introducing-dots/

3.

https://cursor.com/docs/grok-bot

4.

https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/

5.

https://github.com/NousResearch/hermes-agent

6.

https://www.carriermanagement.com/news/2026/09/09/291788.htm

7.

https://docs.openclaw.ai/gateway/configuration-reference

8.

https://docs.openclaw.ai/help/faq-first-run/providers-and-hosting

9.

https://hermes-agent.nousresearch.com/docs/user-guide/security

10.

https://docs.openclaw.ai/gateway/heartbeat

11.

https://www.theverge.com/ai-artificial-intelligence/907074/anthropic-openclaw-claude-subscription-ban

12.

OpenClaw vs. Hermes Agent: The race to build AI assistants that never forget - The New Stack

The New Stack

Berk Kalelioğlu

Yapay Zeka Araştırmacısı

Takip Et

Berk, AIMultiple'ın benchmark ekibinde Yapay Zeka Araştırmacısıdır; etmen tabanlı yapay zeka, makine öğrenimi ve büyük ve küçük dil modelleri (LLM'ler ve SLM'ler) üzerine odaklanmaktadır.

Tam Profili Görüntüle

Yorum yapan ilk kişi olun

E-posta adresiniz yayınlanmayacak. Tüm alanlar gereklidir. Yorumlar orijinal dilinde bırakılır.

Ad

E-posta Adresi

Yorum

0/450

Yorumu Gönder