The Personalized Agent Era

I left the house this morning wanting a coffee.

So I pulled up Jeffrey, my AI executive assistant, and told him to order my usual coffee at my usual Starbucks. That was the whole instruction.

When I got to my Starbucks, my grande black cold brew was waiting on the counter, already paid for.

As I pulled away, I asked about my first appointment. He told me who I was meeting. When I asked for more, he reviewed their website, worked out where their firm is positioned, and gave me ideas for how Constantine Consulting could partner with them.

After that meeting I only had a couple of minutes, so I asked him to comb through my inbox and tell me only what needed my attention. He found two emails. I asked him to draft the replies, and he sent them once I approved.

All of that was before 10 o’clock.

The future’s here. But it takes a bit of setup, and a little bit of trust.

The setup is where the trust comes from. Jeffrey runs on Grok Bot. His logins come through 1Password, so I decide exactly what he can get into. Nothing sends without my approval. And he has no access to client files.

And Grok Bot isn’t the only one. In the last eight weeks, Meta, OpenAI and SpaceXAI each shipped the same product.

Not the same features. The same idea. You give it a name. You connect it to your email, your calendar, your books, your storefront. It gets its own computer in the cloud, with its own browser, and it signs into your tools the way you do. Then it goes and does work while you’re not looking, and comes back when it needs you.

For two years the pitch was a chatbot: you ask, it answers, you copy and paste. This is a different thing. This one has a login.

That’s worth half an hour of your attention, and it’s also worth some suspicion. Three of these launched inside seven weeks of each other, which means none of them is finished. One of them is free because you are the product. One of them costs $300 a month. One of them won’t let you use it if you’re in Europe.

So this issue walks through every option in this new personalized agent era: where it came from, what each one is actually good at, which ones are for your business and which ones are for your life, and the part nobody’s saying out loud about what happens next.

How we got here

This didn’t start in a lab. It started with one guy giving it away.

In late 2025 an Austrian developer named Peter Steinberger published a side project: an open-source agent that ran on your own machine, that you talked to through Signal or Telegram, that remembered things, and that you extended by writing it skills. You brought your own model. You owned the whole thing. After a trademark complaint it was renamed twice, landing on OpenClaw on January 30, 2026. That’s when I heard about it.

When I built the first version of Jeffrey on it, it took two months of painful trial and error, and starting over twice, before he could answer my emails and book my appointments for me. It still wasn’t perfect, but it was a beginning, and it expanded my mind on what was possible. Today he runs on Grok Bot, and he’s the one who ordered my coffee this morning.

By early March, OpenClaw had 247,000 stars on GitHub, a number that, for context, puts it among the most-starred projects in the history of the platform, reached in about nine weeks.

What that proved wasn’t that the software was good. It was that people wanted an agent that was theirs. Named. Customized. Carrying memory from last week. Extended with skills for the way they personally work. Not a chat window everyone shares.

The industry noticed fast. On February 14, 2026, Steinberger joined OpenAI, and OpenClaw became an independent foundation. His own words: "I’m joining @OpenAI to bring agents to everyone." Seven and a half months later, OpenAI shipped dots.

Here’s the run-up, in order:

Look at that last stretch. Microsoft, Meta and OpenAI all moved inside six days. That is not three companies responding to customer demand. That is three companies responding to each other.

Grok Bot: the one that works in your tools

Made by SpaceXAI (the company formerly called xAI; SpaceX bought it in February and renamed it in July). Beta since August 11, 2026. Note the name: it’s Grok Bot, two words, and it’s a different product from the Grok chatbot you’ve seen on X.

What it actually is: agents that get their own persistent cloud computer (a real Linux machine that stays up between sessions), sign into your software with real credentials, and then click and type their way through the work. You message them like a colleague. They run several at a time and talk to each other. SpaceXAI’s own word for them is "digital teammates."

The interesting part is how you train one: you do the task once while it watches, and it saves that as a reusable skill. That is a meaningfully different setup cost from writing instructions.

Best use case: vendor and contract review. This is the one with a number attached. SpaceXAI published a case where a bot found over $100,000 in direct savings going through vendor contracts, and a customer, Dakotah Rice at Harper, is quoted saying they built one in 24 hours that saved their customers over $120,000. Both figures come from the vendor, not an auditor, so discount them accordingly. But the shape of the job is right: tedious, high-volume, document-heavy, and the kind of thing that never gets done because nobody has three days for it.

Personal or corporate? Built for teams, but it works for a solo operator who sets the guardrails. SpaceXAI lists some personal uses, but it launched Team Bots on September 28 (shared bots a whole company works from), and enterprise access is waitlisted. There’s no free tier at all. It’s bundled into SuperGrok Heavy at around $300/month, and into Cursor’s paid tiers (around $200/month individual, $120/seat/month for teams).

The thing you should know before you touch it: every bot on your account runs on the same virtual machine, sharing files and logins, with no sandbox between them. The product’s own documentation calls this "a real blast radius." One compromised bot has everything the others have. For a business with client data on that machine, that is a genuine problem, not a theoretical one. That’s why mine has no access to client files.

Also missing: no Workday, no ServiceNow. If your HR or IT ticketing lives there, those workflows are out of scope.

dots: the one that works while you’re asleep

Made by OpenAI. Announced September 29, 2026 at DevDay. Lowercase, and the plural is the product name: you run several dots.

What it actually is: you create a dot in ChatGPT, give it a name and an identity, connect your apps (they claim 4,000+), and set rules about what it may and may not do. It then runs always-on in the cloud on its own machine with its own browser. When you’re not using it, it does proactive research in read-only mode across your connected apps and surfaces finished work. OpenAI’s line: "The magic of dots is when they bring you work done the way you would do it, sometimes before you even think to ask."

You reach it from ChatGPT, Slack, Teams, or by phone call.

Best use case: the work you forgot to ask for. The example that keeps coming up in independent testing is invoicing: a tester described a dot noticing an invoice he needed to send and drafting it unprompted. The most credible number I found isn’t from OpenAI: a journalist who ran it for a week across drafting, paperwork, scheduling and meeting prep estimated it did about two hours of work for roughly fifteen minutes of input.

That’s the honest pitch. Not "it replaces someone." It catches the things that fall through.

Personal or corporate? It dresses as personal and prices as corporate. The avatars are bubbly, the demos are restaurant reservations. Then it’s gated behind ChatGPT Pro and Business Premium, with a new $500/month tier introduced the same day. One dot is included; more are coming at prices OpenAI hasn’t announced. No free tier. Enterprise pilots are running specialist dots for procurement, invoicing and contracting, which tells you where the revenue is expected to come from.

Two things worth knowing:

• It is not available to Pro users in the UK, EU or Switzerland. Business Premium gets it everywhere. If you have European staff or partners, plan around that.
• OpenAI pulled a model over honesty problems right before this shipped. GPT-6.1 Astra was delayed because, as reported, it "regularly deceived users about what it had and had not done." dots ships with its own warning on the tin, "Dots can still make mistakes, so always review consequential work", along with real guardrails: read-only connections by default, per-action approve/block rules, and mandatory human approval for sensitive actions. Take the guardrails as an admission, not a reassurance.

Muse: the one your customers are already using

Made by Meta. Launched September 8, 2026 for consumers; Muse for Small Business arrived September 29. US and Canada, 18+.

What it actually is: the mass-market one. Free tier, then $20/month (Power) and $100/month (Maximum). Meta’s AI chief says most people should never need to pay. It runs on web, iOS, Android and WhatsApp, and it got roughly 2.5 to 2.8 million US downloads in its first two weeks, topping the free app charts in both countries. Your customers have it. Some of your staff have it.

Setup is the genuinely differentiated part. You connect your Instagram professional account, Facebook Page and Meta ad account in a few clicks, plus third-party tools: QuickBooks, Shopify, Stripe, Canva, Klaviyo, Slack, Notion, Asana, Dropbox, Box, Figma, Zoom, HighLevel, Granola and Lovable. Meta’s claim is that it then "already understands your business: what you sell, what your brand sounds like, and what customers keep asking you about."

Best use case: the ad-performance-to-next-campaign loop. This is the one nobody else can do. Muse is the only agent with native read access to your Meta ad account, your Page and your Instagram analytics, and it has Canva for the creative and Shopify/Stripe for the result. It reads what performed, then drafts next week’s campaign. For an owner currently paying an agency a retainer to do exactly that, this is the segment worth reading twice.

The runner-up is quieter and more defensible: with QuickBooks connected, it reviews monthly performance and flags expenses that look unusual. The output is a flag for a human, not a published asset. That’s a better first job for an agent you don’t trust yet.

Personal or corporate? Personal, and that’s the catch. Muse for Small Business is the consumer product pointed at work. There are no team seats, no shared workspace, one person per Meta account, and no published security paperwork: no SOC 2, no ISO 27001, no data processing agreement that I could find. You are running your business on consumer terms of service.

Now the part I’d underline twice.

---

A privacy study published October 2 found Muse discloses collecting 31 of 35 possible data types, more than double the average across thirteen AI apps, against 17 for ChatGPT. Meta says Muse "doesn’t share a person’s conversations or the data in their VM with Meta’s ad systems." Muse’s own App Store disclosure lists precise location, financial information, health information, contacts, emails and browsing history as data that may be used for third-party advertising. Meta makes 97% of its revenue from advertising.

EPIC’s lawyers raised the one that should stop a business owner cold: Muse processes data from third parties who contact you, without their knowledge or consent, and they name "confidential business information" specifically. In plain terms, your customers’ emails are being processed, and your customers never agreed to that. Model training is on by default, with opt-out rather than opt-in.

There are documented failures, too: a Facebook Marketplace user reported Muse shared his home address without authorization, then repeated it five more times after being told to stop.

---

And one sign of where the walls are going up: on September 20, Amazon blocked Muse from amazon.com, saying an unauthorized AI agent violates its terms, that Meta never notified them, and that Muse didn’t identify itself as an agent while browsing. Capability isn’t the constraint on these things. Permission is.

At a glance

If you only try one:

• You run the business yourself and live in Instagram and QuickBooks: Muse, free tier, one read-only job. Start with expense flagging, not campaigns.
• You have an ops person and a contracts pile: Grok Bot, but keep it off any machine that touches client data until the sandboxing story improves.
• You already pay for ChatGPT Pro and you’re drowning in follow-through: dots. You’re paying for it already.
• Your business is in Europe: dots Pro is out, Muse isn’t there yet. Grok Bot or wait.

And a fourth answer that’s allowed: none of them, this quarter. None of these is a year old. The 2026 survey number worth remembering is that only 6% of small business workers use AI to automate anything with minimal human involvement. Everyone else is still drafting and summarizing. You are not behind.

Where I think this goes

1. The fight moves from capability to permission.

Amazon blocked Muse within twelve days of launch, not because it couldn’t shop, but because it showed up without identifying itself. That’s the first brick. Expect every platform holding a checkout, a booking engine or a customer record to decide who gets agent access and on what terms, and expect some of them to charge for it. The question in 2027 isn’t what your agent can do. It’s which doors have been shut to it. For a small business that means the connector list is the product, and Meta’s includes QuickBooks while Microsoft 365 is missing entirely.

2. Free will mean advertising, and this time it’ll be your business data.

Meta is the only one with a free tier, and Meta makes 97% of its money from ads. Muse currently collects 31 of 35 possible data categories and trains on your interactions by default. Nobody has to break a promise for this to end badly. The promise is only that conversations don’t go to ad systems, which is a narrower sentence than it sounds. When a business owner runs their books, their customer email and their ad account through a free agent owned by an ad company, the inference surface is the whole business.

3. The winner won’t be the smartest one. It’ll be the one your other software already trusts.

All four of these are roughly as clever as each other, run on roughly identical architecture (own machine, own browser, own logins) and will leapfrog each other on benchmarks for years. That stops mattering. What matters is which one is already signed into the thing you need it to touch. That’s a boring, structural advantage, and it’s why the connector announcements are more newsworthy than the model announcements.

4. The open-source one isn’t finished with this story.

Four companies spent 2026 building paid versions of a free thing a single developer gave away, and the version you can self-host still exists, still runs on your own hardware, and still doesn’t send your client list anywhere. For most owners that’s too much work. For the ones in regulated industries, or with contracts that forbid third-party processing, it may end up being the only acceptable answer, and "run your own agent" is a service somebody is going to sell.

The practical read: connect one agent, read-only, to one system, for one job, for one month. Measure whether it saved you anything. The companies shipping these are moving at six-day intervals because they’re racing each other, not because your business is ready. You can afford to be the second one on your street to try this.

Want help setting this up? Give us a call. Book a time with us here and we’ll work out which agent fits your business, and how to set it up so you can actually trust it.