AI Agents Just Got Phone Numbers, Wallets and a Seat in the Group Chat

I stumbled on an interesting X post this week on AI agents from McBain (@heyneighbor).

I’ve been building conversational AI since 2018, back when “chatbot” mostly meant a decision tree in Facebook Messenger. I’ve seen plenty of hype cycles. What stood out this week was a pattern across launches: the agent is leaving the chat window. It’s getting a phone number, an inbox, a payment card and a seat in your group chat. Once software has its own contact details, you start treating it like a team member.

If you run a business, that changes who answers your phone, who touches your money and who sits in your team’s threads.

Here’s how this post is laid out:

The quick roundup. 18 ai agents you can scan it in a minute. Skip ahead if you just want my take.
What it means for your business. Five shifts I’m watching, each with one thing you can do this week.
Where this leaves us. Why I’m hopeful, and one question I want you to answer.

Part 1: The quick roundup

Here’s 18 personal AI agents grouped so you can scan it in a minute.

Big labs

ChatGPT dots (OpenAI): You can now create a dot from the iOS/Android app. Dots can start Codex work and edit Scheduled Tasks. The Register’s test dot burned 300K+ tokens in a day, unprompted.
Grok Bot (SpaceXAI): Runs Shopify stores. It also got its own email address, X search with no connector, Figma/Notion/Canva in chat, and payments via X Money or Stripe Link. It’s in every Cybercab and the Model Y fleet, and Grok Voice 3.0 now runs its voice mode.
Muse (from Alexandr Wang): Announced Muse Wonder and Wonder Full. Wonder uses the Meta Model API plus Muse Code. Wonder is headed to the personal assistant but isn’t confirmed live there yet.
Google and Anthropic: Mentioned only for being absent. McBain asks when they’ll ship agents of their own.

Open source / self-host

OpenClaw: v2026.10.1 added installed-skill discovery, opt-in MCP Apps beside chat, external/R2 full backups and Mac setup fixes, with 402 contributors. Follow-up releases added Muse Wonder support.
Hermes Agent (Nous Research): A big update made it easier to get started, connect your apps and run AI models on your own computer with one click. Two quick fixes followed, including one for a broken update.
NanoClaw: Its first official stable release, with clear labels separating stable versions from early test versions. It also shipped an important security fix for the account logins it uses on your behalf.

Proactive assistants

Eney (MacPaw): Launched on Mac and iPhone. It spots upcoming calls and tasks, prepares next steps and asks before acting.
Hark Pro (Hark Labs): Action Buttons rank tasks it notices in your connected accounts and run them in one tap. Its CEO, Brett Adcock, took a polygraph to back its “we never sell your data” pledge.
Ambi (Mistlabs): An ambient assistant, now out of beta on Apple devices and backed by $26M. It learns from conversations and surfaces suggestions. theCUBE reported the news. The standalone wearable is still to come.

Group-chat agents

Rene (Rene Labs): Now works in group chats. It builds itineraries, calls restaurants, designs invites, buys supplies, makes budget sheets and sends reminders.
Genspark Claw: Now works from iMessage, WhatsApp, Slack or LINE. It added Muse Wonder and Gemini 4 Argon, and it promised credits after an outage it blamed on a Microsoft service issue.

Calls and payments

Lindy: Learns a task by watching you do it once, then runs it on a schedule. New Phone Calls feature: call it, have it call you, or use it as a receptionist. A third-party tracker shows its plans reshuffled to Free / Team / Enterprise.
Pally: Now completely free. That includes the top models, all app connections, unlimited automations, its own phone number and email, and calls and web purchases on your behalf.
Underdog: On-device agent payments with Stripe Link, using a one-time card that waits for your OK. It also added local image and video generation and a Windows app, and its Saluki 27B model passed 300K downloads in its first week.
Stripe: Its Link payments sit behind both Underdog and Grok Bot.

Platforms, partnerships and funding

Perplexity Computer: Explained its Projects filesystem as a Git forge for agents, added GPT-6 Astra and SOL 6.1 for Max subscribers, and open-sourced SPACE, its task sandbox, under Apache 2.0.
Manus: Partnering with OPPO on Agent Mode for AI Mind Pilot, planned to debut on the Find X10 series with ColorOS 17. Markets are still to be announced.
Persona: From Cal AI co-founder Zach Yadegari. TechCrunch reports $10M raised, led by Vine Ventures. It has a free iMessage beta, and a $179 wearable band is expected in December.
Town: Plans now get “about twice the work, for the same price.” It was named to Forbes’ Agentic 20, and team features are coming.

Part 2: Five shifts worth your attention

Now for what I think matters, and what I’d do about each one.

1. Every agent can pick up the phone now

Lindy can answer your phones as a receptionist. Pally gives every user an agent with its own phone number and email, for free. Rene calls the restaurant for your group.

I spent years in enterprise conversational AI, at knowbl and at VAgents, where we worked on systems handling 500M+ calls. Voice was always the hardest, most expensive part to get right. Seeing it bundled into consumer and small-business agents, sometimes at no cost, is a big deal. A capability that used to need an enterprise budget is becoming a checkbox.

What to do: Look at your missed calls. If after-hours calls or overflow are costing you leads, this is the cheapest it has ever been to test an AI receptionist. Start with one narrow job, like booking or basic FAQs, and listen to the recordings before you widen its scope.

2. Agents can make purchases now

Look at how the payment features were built. Underdog’s agent uses a one-time card that waits for your OK. MacPaw’s Eney prepares next steps and asks before acting. Grok Bot pays through X Money or Stripe Link.

Now look at the warning sign. The Register’s test ChatGPT dot burned 300K+ tokens in a day without being asked. An agent that works on its own also spends on its own.

Trust gets built in small, reliable moments, and agents work the same way. That’s also why Hark’s CEO took a polygraph to back a privacy pledge. Privacy has become something companies sell on.

What to do: Before you give any agent a card, an inbox or your CRM, write down three things. What can it do without asking? What needs your approval? What’s the spending cap? If a vendor can’t support those answers, it isn’t ready for your business.

3. Agents can hang out in group Channels

Rene works inside group chats. Genspark answers you right where you message it, on iMessage, WhatsApp, Slack or LINE. Town says team features are coming.

This is the part I find most exciting. Most coordination doesn’t happen in one person’s productivity app. It happens in threads: the sales team’s Slack channel, the client’s WhatsApp group, the family chat planning a trip. An agent in that thread serves the whole group.

What to do: Pick one recurring coordination job your team handles in chat, like event logistics, follow-up reminders or a weekly status roundup. That’s your first shared-agent pilot.

4. Free agents are getting good!

Not every AI agent is a service you sign up for. Some are free, open-source programs: the code is public, and you run them on your own computer or server. Three of them, OpenClaw, Hermes Agent and NanoClaw, all put out updates this week.

The updates are quiet, and quiet is good news here. OpenClaw added full backups, so you can keep a complete copy of your setup somewhere safe. NanoClaw fixed a security problem in how it handles the account logins it uses on your behalf, and it now clearly labels which updates are stable and which are early test versions. Hermes Agent shipped a quick repair after one of its updates went out broken. Backups, security fixes and fast repairs are what software looks like once real people depend on it.

I have a personal stake here. I’m building my own team of AI agents on Hermes Agent, with shared notes in Obsidian so every agent works from the same information. This week’s Hermes update caught my eye: it can now connect to the apps you already use, and it can run an AI model right on your own computer with one click. So did the speed of that repair. When an agent holds the keys to your accounts, treat every update like a security notice: read what changed and install the fixes.

A short exchange in the replies summed up the trade-off. McBain pointed out that you can text OpenClaw on iMessage. A reader pushed back: doesn’t each user have to set it up themselves, rather than just texting a ready-made service? McBain’s reply: “Oh yeah.” Do-it-yourself software gives you control and ownership, and in return you do the setup.

What to do: If you have someone technical on hand and you care about owning your data and the way your work gets done, running your own agent is worth a look. If you don’t, the ready-made services above are improving fast. Either way, your agents need one up-to-date place that holds the basics: your offers, pricing, policies and tone. The agent is only as good as the information you give it.

5. Measure your Agent in dollars, not minutes

Zach Yadegari, founder of Persona, said something interesting:

“The average consumer actually does not value their time… What people do care about is saving or making money.”

I think he’s onto something.

Almost every agent is pitched as a time-saver, and “saves you 10 hours a week” sounds great. But hours are easy to lose track of. Dollars aren’t.

What to do: When you test any of the agents above, track the result in money: leads your AI receptionist booked, deals your shopping agent found, costs you cut. If you sell AI services yourself, use the same lens in your pitch and lead with the dollar outcome.

Part 3: Where this leaves us

I used to present a workshop called Finding Digital Balance, and I still believe the goal of technology is to give us back our lives, not to take them over.

This week’s pattern gives me hope.

The best-designed agents act first and ask before anything important.

That’s the right balance: let the agent do the legwork, and keep a human in charge of the money, the relationships and the judgment calls.

The personal agent is becoming a participant in your business. As more of the internet gets automated, the people you trust and the rules you set become your edge. The winners will be the businesses that set clear rules for their agents, however many they run.

My question for you: If you could hand one job to an agent with its own phone number and a card that waits for your OK, what would it be?

Create a Great Day!
Arvell Craig

About the Author

Arvell Craig is Director of AI at BotBuilders, an Inc. 5000 company helping small businesses implement AI-powered systems. He is a keynote speaker, AI strategist, and author of The AI Strategy Playbook. He speaks and writes on AI fluency, agentic adoption, and what it takes to build organizations that are genuinely ready for an AI-disrupted world.

Book a Call
Connect on LinkedIn