How the Big AI Companies Are Building Agents

By Steve Macfarlane, CEO of Syfre

The other day, in a planning session with my senior developer at Syfre, the conversation landed on AI agents. The term is everywhere right now, on the podcasts I listen to and at pretty much every networking event I go to, and I’ve noticed that most people, even some very tech-savvy ones, are still pretty vague on what it means.

I do wonder whether agents are a real shift, or just a new label for something all the AI companies have been organically heading towards anyway. I also wonder how much of it is marketing hype. But ultimately I don’t think it is. My gut tells me that agents are a real shift, and one that’s going to change how a lot of businesses operate.

So far most of us have used AI chatbots to produce things for us, like a draft email or a summary of a long report. And however good the result, it stops at the words on the screen. Sending the email or acting on the report is still our job. An agent, on the other hand, moves that relationship towards delegation, where you hand it the job itself and it works out the steps. It’s the difference between asking a trusted colleague to help you write a briefing for a client meeting, and handing them the meeting itself. Like, “we’re seeing the client next Thursday. Find a time that works for everyone, book the room, pull together what they’ve been up to lately, and have a briefing in my inbox the night before.”

Every major AI company seems to be heading in that direction, they’ve all just started from very different places, which complicates the topic. Meta is aiming squarely at everyday life, the errands and the family admin. Anthropic has gone after complex professional work, and most of the others sit somewhere in between.

I’ve been a fair way down this AI agent rabbit hole lately, and a good anchor anecdote about it all is the origin story. It all starts with a developer in Vienna named Peter Steinberger. In May last year, Steinberger had an idea for a personal AI assistant, but he decided to wait for one of the major tech companies to build it. By November he got impatient because nobody had built what he had in mind, so he built it himself. I can relate to the feeling, it’s hard to have patience when everything in tech seems to be moving at such an insane pace right now. I just heard on a podcast that the big AI labs are now releasing a new frontier model every ten days or so, which sounded like an exaggeration until I looked it up. ARK Invest’s numbers put it at roughly every 11 days in the first part of this year, down from 17 days in 2025. I know it always feels like life is moving fast, but at the moment that sentiment is quite literal.

The first version of Steinberger’s idea, after he took matters into his own hands, took him only about an hour to build. It took the messages he sent on WhatsApp and passed them to an AI coding tool on his computer, which sent the answers back, and he called it WhatsApp Relay.

He’d already built a successful software company and sold most of his stake in it, so this wasn’t a full business plan, it was more of a passion project. Again, I can relate. It’s a challenge as a developer not to jump all-in on every exciting new idea. In Steinberger’s case, what he wanted turned out to be pretty straightforward and easy to build, which is even harder to resist. He wanted something that lived on his own computer, took instructions through the messaging apps he already used, and got on with jobs while he did other things. After some tweaks he renamed it Clawdbot and was happy enough to put the code on GitHub for anyone who wanted it.

It turns out that a lot of people wanted it. After a couple more name changes it became OpenClaw, and within nine months it had more than 387,000 stars on GitHub, which is the developer equivalent of a standing ovation. By February, both Mark Zuckerberg and Sam Altman had made him job offers. Zuckerberg’s offer came over WhatsApp, and by Steinberger’s account the two of them spent ten minutes arguing about which AI model was better. He ultimately took the job at OpenAI, and the project moved to an independent foundation.

Then the big companies arrived. Google went first with Gemini Spark in May of this year, and its advantage is a pretty obvious one. It already has your email, so it comes connected to Gmail and the rest of Google’s apps from day one, which is a head start the others will struggle to match. But the three I want to spend time on here are the ones I think tell us the most about where this is all heading. xAI released Grok Bot in August. Meta’s Muse followed in early September, and OpenAI launched its own just recently at the end of September, which it named dots. Those three landed within seven weeks of each other, and if you line them up next to what Steinberger built, the resemblance is pretty clear. The basic shape of a personal agent seems to be settled. What separates them now is mostly a question of trust, and in particular of who holds the keys to your accounts.

An easy way to understand the difference between working with chatbots and working with agents is to consider an ordinary job, like planning travel. Ask a chatbot to sort out flights for a conference in Melbourne and you’ll get a decent list of options, but that’s where its involvement ends. Hand the same job to an agent and it compares the fares, picks the one that fits your calendar, checks with you before it pays, files the receipt, and then later messages you if the airline moves the departure time.

An agent keeps working after you’ve closed the app. It comes back to you when something needs a decision, and the rest of the time it just gets on with things mostly autonomously.

It also sticks around to build memory and context. It’s like a good assistant who takes on a steady stream of jobs, anything from moving a meeting to a project that runs for a week. It remembers how you like things done and can run the same report every Monday without being asked. Some will even go looking for loose ends while you’re busy. And it’s still there the following week, with a slightly better sense of how you work.

To do any of this, an agent needs somewhere to work. Every product here gives the agent a computer to work on, with a browser, files, logins and some kind of memory, and every one of them lets you talk to it the way you’d message a colleague. OpenClaw runs on a machine you own, which can be the laptop on your desk. The big companies give each agent a virtual computer in their own data centres, a complete computer that exists as software, so the agent can log into your accounts and browse the way you would while your own laptop stays closed.

Meta’s Muse is the most consumer-minded of the group. Meta calls it the world’s first personal AI agent built for everyone, which is a bold claim given OpenClaw got there first, though to be fair to Meta, “for everyone” is the important part. Muse lives inside WhatsApp and its own app, and it’s designed for people who have never typed a command into anything. You message it the way you’d message a friend or colleague and it goes off and handles things. Meta’s own examples are mostly domestic. Save a recipe reel on Instagram and Muse can turn it into a shopping list. Mention a holiday and it can research the options and line up a booking for you to approve.

Meta has put a lot of its effort into security, and it has the most careful design of the group. Each person’s Muse runs on its own sealed-off virtual computer. A second agent called Sentinel sits beside it and has to approve anything that reaches the internet or touches a connected app. Your passwords are kept somewhere the agent can’t see, so it works with stand-in tokens, and when it buys something it pays with a virtual card number that only works once. Meta also released a small-business version at the end of September that plugs into tools like Slack, Asana and Canva, which is a fair hint about where this is all heading. (The catch for now is that Muse is only in the US and Canada, so Australians are watching from the sidelines.)

xAI has pointed Grok Bot at work instead, and describes it as AI teammates you can give real work to. You can set up several named bots, each with its own responsibility, and they share a single cloud computer with a screen each. They work through the tools you connect and come back to you when they need input.

A feature I find really interesting on Grok Bot is called teach-a-task. You press a button, the bot starts recording, and you do the job yourself while it watches. When you stop, it turns what it saw into a skill it can repeat on its own. One reviewer taught a bot to pull newsletter statistics from a publishing platform in a single recording, and it’s produced the report on request ever since.

Almost every business has jobs that belong to just a single person. Someone logs into a couple of supplier portals on the first Monday of the month, for example, copies the numbers into a spreadsheet and emails it to the same four people. Automating that kind of job has usually meant a developer and a project code. With teach-a-task, the person who already does the job records it once, in ten minutes or less.

Getting hold of Grok Bot is easier now than it was at launch. It now comes with xAI’s paid plans from US$30 a month, and with Cursor’s paid plans too, the coding company SpaceX bought in August for $60 billion. (xAI is part of SpaceX now as well. Busy year for Musk.)

OpenAI’s dots are built for the more complex jobs. You give a dot a goal, connect the apps it needs, and it keeps making progress over days, coming back when it needs direction or approval. Each dot runs on its own cloud computer and can reach more than 4,000 apps through OpenAI’s plugins. You can message it in ChatGPT, or in Slack and Teams, so it can sit in the same channels as the rest of your team. When you’re not working with it, it does background research using read-only tools, so it can look around without being able to send or change anything.

The part I like most with dots is the rulebook. OpenAI calls them Custom Rules, and they’re where you decide how much a dot can do before it has to check with you. Some things are off limits regardless. Changing a password, for instance, always stays with the human. The rules matter because there’s a big difference between an agent that drafts an invoice and one that sends it to your customer. You can check a draft before anyone else sees it.

One of OpenAI’s early testers had a dot notice an invoice that had never gone out. It prepared the invoice, asked, and sent it once the tester said yes. Most of what these agents will do in a business looks like that, the small dropped thread someone meant to get to and never did. Dots are limited to the more expensive Pro and Business Premium plans for now.

OpenClaw is the one you run yourself. The software is free (you pay for whichever AI model it uses), it lives on your own computer or a server you set up, and you choose the model behind it, whether that’s OpenAI’s, Anthropic’s or someone else’s. You talk to it through WhatsApp, Telegram, Slack or iMessage. Its memory is a folder of plain text files you can open and read, including one called SOUL.md that the agent reads at the start of every session to remember who it is, which is pretty cool.

The trade-off is that the safety work Meta does for you with Muse becomes your job here. You decide what the agent can get into and which add-ons to trust.

The person OpenClaw suits is a developer or a seasoned tinkerer, someone who enjoys the plumbing and wants an agent on a spare computer that sends them a briefing on Telegram every morning. It probably isn’t the right first agent for your CFO.

Anthropic, which makes Claude, is the odd one out, because it hasn’t launched a personal assistant at all. Its agents are built around work, and they operate inside the places work is already happening. Claude Code, which writes and fixes software, gets most of the focus, but the closer comparison to everything above is Claude Cowork, a desktop app that takes on larger jobs involving your own files and documents. In Slack, anyone on a team can tag Claude into a conversation and hand it a task. Underneath all of it sits what Anthropic calls Skills, which are packages of instructions, reference material and sometimes code that teach Claude how a particular business does a particular job, like its monthly reporting or its brand guidelines.

It’s a different version of the same idea. An agent that can click through software still needs to know how your organisation likes things done, and Anthropic has put its effort into that part.

None of this is risk-free, of course. The one risk that security people talk about most is prompt injection, which is roughly phishing aimed at the agent. Someone hides instructions in an email or a web page, hoping the agent will read them and follow along. It’s a well-known problem, which is part of why Sentinel and the rules in dots about what has to be checked first exist at all. 

The practical answer is the same one you’d use with a new hire. Give it jobs where a mistake is easily fixable to begin with, like a report someone reviews before it goes anywhere, and widen what it can do as it earns your trust. Moving money and making commitments in the business’s name can come later.

If it helps to picture them as people, Muse is a concierge in a well-guarded building. It runs your errands and never holds your keys, only a card that works once. Grok Bot is the new hire who sits beside you for a morning and watches how you do things. A dot is an executive assistant who turns up on day one with a written list of what they’re allowed to sign. Claude is the specialist who already has a desk in your office and a copy of the procedures manual. And OpenClaw is the gun freelancer you hire directly, with no agency in between. You set the terms and can read every note they keep, and you can check the references of anyone they bring in to help.

There’s an obvious way to read all of this. An agent that handles the reporting and chases the invoices looks a lot like a role you no longer need to fill with a person, and for some businesses that maths will work, up to a point. But I think that’s the smaller move. The bolder move is to give the agent to your best account manager instead, and see what they do with the hours they get back. Someone just has to be intentional about deciding what that time is for, or it fills straight back up with admin.

Which is more or less where Steinberger started. He wanted an assistant to take the work he didn’t need to be doing himself, and he wanted it running on a computer he controlled. He works at OpenAI now, one of the companies he’d been waiting on. The project he started that November belongs to a foundation, and it still runs on whichever computer you decide to trust it with.

—

Steve Macfarlane is the CEO of Syfre, an Australian AI software studio based in Brisbane.