Autonomous Agents Are a Skill Now. Seven Options, Compared
November 6, 2026 | AI Tools & Business
Between August 11 and October 21, SpaceXAI, Meta, OpenAI and Anthropic each shipped an autonomous agent that runs on its own cloud computer and works inside your apps. They joined two open-source projects, OpenClaw and Hermes, and an invite-only startup called Instinct. Picking one matters less than learning how to run one, because the skill transfers and the product lineup will not hold still.
Here is the case for learning this now, followed by all seven compared on cost, integrations, ease of use and risk. Prices were checked November 6, 2026.
Why This Is Worth Learning Now
A chatbot answers. An agent acts. It signs into your tools, works while you are away, and comes back with finished work or a mistake already made. That difference changes the job from writing good prompts to supervising a worker.
Employers are already pricing that in. In KPMG's Q1 2026 AI Pulse, 57% of leaders said they expect people to manage and direct AI agents, and 76% named skills gaps as a barrier. Randstad reports that job postings requiring AI agent skills rose 1,587% during 2025. I have not seen a baseline for that number, so read it as direction, not precision.
The other reason is personal. I run three scheduled agents inside PAID LLC, and the expensive lessons had nothing to do with prompts:
- Agents fail by going quiet. Our daily blog ran 21 straight days, then stopped. Nothing broke and nobody decided to end it. We now compare the runs each agent should have made against the runs it logged, so a silent agent shows up as a failure.
- A safety limit you have not tested is a guess. Our daily spending caps were silently off from June 16 to September 19 because a database function they depended on was missing, and they were built to fail open. Nothing warned us for three months.
- The approval line has to be absolute. No agent in this business sends an email, publishes, spends money or contacts a person. Each one drafts and stops, and a person does the rest.
The Seven at a Glance
The ease column is my judgment from each product's documentation and published reviews.
| Agent | Starting cost | Access | Integrations | Ease | Biggest risk |
|---|---|---|---|---|---|
| Grok Bot | $30/mo | Open | Gmail, Slack, Notion, GitHub, browser | Moderate | Can't choose the model |
| Instinct | $0, no published price | Invite or waitlist | Text, calls, Google Workspace | Easiest | Terms put the risk on you |
| Muse | $0, then $20 or $100/mo | Limited locations | Gmail, Calendar, Shopify, QuickBooks | Easy | Handing Meta your inbox |
| Dots | $100/mo (ChatGPT Pro) | Paid plans, some regions excluded | 4,000+ apps, Slack, Teams | Moderate | Usage caps unpublished |
| Claude Duet | $20/mo (Claude Pro) | US waitlist | 6,000+ connectors, no Microsoft 365 yet | Easy | Slow checks, early preview |
| OpenClaw | $0 software, roughly $25 to $60/mo to run | Open source | 10+ chat apps, any model | Hard | Security is your job |
| Hermes | $0 software, roughly $6 to $85/mo to run | Open source | 20+ platforms, any compatible model | Moderate to hard | Security is your job |
The Closed Products
Grok Bot (SpaceXAI, formerly xAI)
Grok Bot signs into apps and operates their screens the way an employee would, so it works where no clean connector exists. You can show it a workflow once and it saves the routine. A Chief of Staff Bot can coordinate specialist Bots.
- Cost: New customers get it through The X Subscription at $30, $100 or $200 a month. The Free and $8 Plus tiers exclude it. When I wrote about Grok Bot on August 31, the cheapest path was $20 through Cursor Pro. Six weeks later, new customers paid $30.
- Pros: Works on Mac, Windows, Linux, iPhone, iPad and Android, and purchases require approval on every spend.
- Cons: You can't choose or pin the model, independent reviewers describe limits that run out fast, and all your Bots share one computer and one set of logins.
Instinct (Spear Street Technology)
Instinct has the lowest friction on this list. You text it on iMessage or WhatsApp, or call it, and it books, negotiates and follows up. It also texts you first.
- Cost: Free during the invite-only beta. No price is published, and the terms allow pricing to change. Whatever it books is billed to you by the merchant.
- Pros: No app to learn. It connects to Google Workspace and uses one-time Stripe cards for purchases.
- Cons: Agreements it makes bind you as if you made them yourself, and Instinct disclaims unintended actions. Model training is on by default, and disconnecting an account does not delete indexed data. One early tester emailed malicious instructions to their own inbox and Instinct followed them, according to a launch-week account compiled by Vellum, a competitor.
I would keep Instinct on personal errands and give it read-only access to any work inbox.
Muse (Meta)
Muse launched September 8 and is the cheapest serious way to start. Zuckerberg says more than 10 million people in the US and Canada have set up an agent.
- Cost: Free with a usage limit. Power is $20 a month for 500M Muse tokens a week, and Maximum is $100 for 3B. Meta does not say how many tasks a token allowance covers.
- Pros: Strong controls for a consumer product: an approval layer called Sentinel, per-app scopes, an audit log, and a credential vault Muse can't read. Small business connectors reportedly added Shopify, QuickBooks, Stripe and Canva.
- Cons: Meta says Muse data stays out of its ad systems, but you are still giving Meta your email and calendar. Meta also describes it as in limited testing and not yet available everywhere.
Dots (OpenAI)
Dots launched September 29 and is the natural fit for a team already living in Microsoft Teams. Each dot runs on GPT-6 Astra in its own cloud computer.
- Cost: The first dot is included with ChatGPT Pro ($100, $200 or $500 a month) and Business Premium ($125 per user monthly, two seat minimum). Plus does not include one. OpenAI has not published the price of extra dots or the usage caps. Work a dot hands to Codex or ChatGPT Work counts separately. Enterprise admins should check their console now, because OpenAI's help page lists November 18 as the date billing starts.
- Pros: More than 4,000 apps through plugins, Custom Rules that allow, gate or block specific actions, and an activity view for watching background work.
- Cons: It carries the highest entry price here, and Pro users in the EEA, Switzerland and the UK are excluded for now. Budgeting is guesswork until the caps are public.
Claude Duet (Anthropic)
Disclosure first: I pay for Claude, and Claude co-wrote this post. Weigh this section accordingly.
Duet launched October 21 as a research preview built for work. It reads, drafts and prepares on its own, then waits for your approval before it sends, pays, signs or deletes.
- Cost: Included at no extra cost with Claude Pro ($20 a month, or $17 billed annually), Max ($100) and Ultra ($200). Background work draws down the same limits as the rest of Claude.
- Pros: Over 6,000 connectors and plugins, including Gmail, Slack, QuickBooks, Stripe, HubSpot and Docusign. It can't trade securities or move money between accounts, even with approval. That is the most conservative default of the seven.
- Cons: US waitlist only, with no Team or Enterprise plans and no Microsoft 365 yet. It checks accounts every few hours, so urgent items wait. You can't watch or take over its browser, and Anthropic itself warns it can misread a thread.
The Open-Source Options
OpenClaw
OpenClaw runs on your own machine or server and reaches you through the chat apps you already use, including WhatsApp, Telegram, Slack, Discord, Signal, iMessage and Teams. It is MIT licensed and works with the model of your choice, including local ones.
- Cost: The software is free and a basic server runs about $4 to $10 a month. Model usage is 90% or more of the real bill. Operator reports collected by Kilo put typical setups at $25 to $60 a month, premium models at $85 to $270, and heavy multi-agent work at $210 or more. An idle agent on a premium model can burn about $5 a day just checking in.
- Pros: No lock-in, full control of your data, and the largest ecosystem here.
- Cons: Its 2026 security record is the worst on this list. CVE-2026-25253 allowed one-click remote code execution until version 2026.1.29. Koi Security found 341 malicious skills among the 2,857 on ClawHub, and Bitdefender's scan put the number near 900. SecurityScorecard found more than 30,000 instances exposed to the internet.
Run OpenClaw on a dedicated machine, keep it patched, and read every skill before you install it.
Hermes Agent (Nous Research)
Hermes is the open-source agent built around learning. It keeps its own memory, recalls past sessions, and writes and improves skills from experience. It is MIT licensed and connects to more than 20 platforms, including Telegram, Slack, WhatsApp, Signal and Teams.
- Cost: Free software. Hostinger, which sells Hermes hosting, estimates $6 to $85 or more a month self-hosted. Nous Portal plans are reported at $0, $20, $100 and $200.
- Pros: The Windows and Mac installer and a single Portal login remove most of the server work, and it runs on Nous Portal, OpenRouter, OpenAI or any compatible endpoint.
- Cons: It has its own CVEs, including a high-severity remote code execution flaw (CVE-2026-71963) in versions 0.18.2 through 0.21.0, fixed by a September commit. Update before you run it, and treat it with the same care as OpenClaw.
Which One to Start With
- You want to learn for free: Muse, on a low-stakes workflow, and accept the Meta trade.
- You run on Google Workspace, QuickBooks or Stripe and already pay for Claude: Join the Duet waitlist.
- Your team lives in Microsoft Teams: Dots. Duet can't reach Microsoft 365 yet.
- You need a screen-driving agent for tools with no connector: Grok Bot at $30.
- You need control of your data or model choice, and someone technical is available: Hermes first, OpenClaw if you need its ecosystem.
- You want a personal errand runner: Instinct, once you have read the terms.
A 30-Day Plan
Whichever you pick, run the same four weeks:
- Week 1, read only. Connect one tool and have the agent summarize something you already review by hand.
- Week 2, drafts. Let it prepare the work. You send everything.
- Week 3, one approved action. Allow a single reversible action with your approval on every run.
- Week 4, measure. Track time to a first useful draft, how much editing remains, and how often you had to step in.
If those numbers improve, widen the scope one permission at a time. Test your spending limit on purpose, and check that the agent actually ran.
The product you choose today will have a different price by spring. The supervision habits are what you keep.
Sources: KPMG, Q1 2026 AI Pulse. Randstad figures via HRKatha. VentureBeat on Grok Bot. GrokMag, Grok Bot pricing and limits. eesel, Instinct review. Carly, Instinct pricing. Vellum, Instinct breakdown. Meta, About Muse subscriptions. eesel, Muse review. CNBC on Claude Duet and Muse. VentureBeat on Dots. OpenAI, Getting started with your dot. Anthropic, Introducing Claude Duet. Anthropic, Get started with Claude Duet. OpenClaw docs. Kilo, OpenClaw cost. DigitalOcean, OpenClaw security. Hermes Agent docs. Hostinger, Hermes cost. OSV, CVE-2026-71963.
Written by Travis Raveling, Founder PAID LLC, co-authored and edited by AI.
About PAID LLC: PAID LLC helps small and mid-size businesses implement AI tools that save time and drive revenue. See our services at paiddev.com/services.