)
If you have read a tech headline in the past month and felt slightly behind, you are in the same position as roughly three quarters of Australian small and medium business owners.
AI is in the news every day. The vocabulary keeps changing. The companies keep shipping things with names that sound like rejected sci-fi novels.
By the time you have a confident grasp of one development, the conversation has moved on.
This article is the first in a short thread for Business News Australia readers. It is deliberately patient, it assumes nothing, and it avoids the jargon that makes most AI coverage exhausting to read.
The aim is simple: by the end of this piece, you will understand what AI agents are, why they are different from the chatbots you have already used, and why three of the four largest technology companies in the world spent the last fortnight in a public race to ship one.
You do not need to act on any of this today. You do need to understand it before you make your next IT spending decision.
First, the thing you already know The AI you have probably used (ChatGPT, Claude, Google Gemini, Microsoft Copilot) is what the industry calls a "chatbot" or, more technically, a "large language model".
You type something. It types something back. The conversation lives inside one window in your browser. When you close the tab, it forgets you. This is useful but limited.
The chatbot can write you an email, but it cannot send the email.
It can analyse your sales numbers, but only the sales numbers you copy and paste into it. It can suggest you reschedule a meeting, but it cannot actually move the meeting in your calendar.
The chatbot is, in a real sense, trapped in the box.
For two years, the chatbot has been the dominant form of AI most people interact with. That is now changing.
What an "AI agent" actually is An AI agent is a chatbot that has been given hands.
It uses the same underlying intelligence (the same large language model) but it can also do things. Open your email. Read it. Respond to it. Look up information on a website. Fill out a form. Pull a row from your accounting software. Add an event to your calendar. Send a message in Slack. Move a file from one folder to another. Run a small piece of code. Then keep going, take the next step, and the step after that, until the task is done.
The conversation analogue is not "I asked the AI a question and it answered". It is closer to "I gave the AI a goal and it figured out the steps and did them, occasionally checking in with me to confirm the bigger decisions".
That is a meaningfully different relationship with the technology.
A useful concrete example: an Australian SME could say to an AI agent, "every Friday afternoon, look at our Xero account, identify any invoices that are more than 30 days overdue, draft polite follow-up emails to the relevant clients, and put them in my drafts folder for me to review on Monday".
The agent does the looking, the identifying, the drafting and the placing. The owner does the reviewing and the sending.
That workflow used to belong to a junior bookkeeper. It now belongs, increasingly, to software.
Why this is happening now
Three things have changed in the last 12 months that, taken together, made the agent shift possible.
The AI models themselves got smart enough to plan. A 2023-era chatbot could write you an email. A 2026-era model can hold a multi-step plan in its head, recognise when it has hit a problem, change course, and recover.
That is what is required to do real work without a human looking over its shoulder for every keystroke.
The models also got cheaper to run. The cost of having an AI think through a hard problem has dropped roughly 90 per cent in two years.
That economics shift is what makes "leave the agent running for an hour and let it work" affordable rather than ruinous.
And the software that connects the AI to the rest of your computer (the part that actually lets it open Xero, read your inbox, click around your CRM) has matured rapidly.
This piece of software has a name that you will hear more often this year: the orchestration layer, sometimes called the agent harness.
It is the layer that sits between the AI's thinking and the actual doing.
Think of it as the chassis of the car. The engine is the AI model. The chassis is what turns the engine's power into something that actually goes somewhere.
Where OpenClaw fits in
You may have seen the name OpenClaw in tech headlines over the past few months. It is currently the most-talked-about piece of AI software in the world.
OpenClaw is, in plain terms, an open-source orchestration layer.
It is the chassis. Not the engine. It does not have its own AI model inside it; instead, it lets you plug in whichever AI you want to use as the brain (Claude, ChatGPT, Gemini, or one of several free options that run on your own hardware).
What makes OpenClaw notable is three things.
It is free. Anyone can download it. It runs on your own computer rather than in someone else's cloud.
That means your data, your business documents, your client information, and your workflow history stay on a machine you own.
And it is genuinely open. The code is public, anyone can inspect it, and no single company controls it.
OpenClaw was started as a weekend hobby project by an Austrian developer in November 2025.
By March this year, it had become the most-starred non-aggregator software project in the history of GitHub, the platform where the world's developers share code.
It got there in 60 days. The previous record holder, a piece of web infrastructure called React, took over a decade.
That growth rate is what tipped off the largest technology companies that something significant was happening.
The fact that millions of developers had quietly adopted an open-source agent that they could run on their own hardware, with their own data, was a strategic problem for any company whose business model depends on you running your AI workloads in their cloud.
The new race
Over the last fortnight, this dynamic has produced a flurry of news that, from the outside, can look like four separate stories. It is one story.
Anthropic, the maker of Claude, restricted how its software can be used with third-party tools like OpenClaw.
Users now pay per use rather than via flat subscriptions when they connect Claude to outside agents.
OpenAI, the maker of ChatGPT, did the opposite.
The company hired the founder of OpenClaw and publicly endorsed using ChatGPT subscriptions inside whichever third-party agent users prefer.
Meta, the parent of Facebook, Instagram, and WhatsApp, was reported to be building its own consumer version of OpenClaw, codenamed Hatch, designed for non-technical users.
Google was reported to be testing its own equivalent, codenamed Remy, which industry sources suggest could be previewed at the company's annual I/O developer conference on 19 May.
And Apple, somewhat to its own surprise, sold out of Mac Minis. The company's chief executive Tim Cook told analysts on a recent earnings call that Mac Minis and Mac Studios are running short worldwide because of demand from people running AI agents on their own hardware.
The shortage has reached Australia. I have spent the past fortnight visiting Apple flagship stores, JB Hi-Fi, and Harvey Norman trying to buy one. None had stock in any meaningful configuration.
Why an Australian business owner should care
Three reasons, briefly.
The first is that AI agents are about to become the default interface to a lot of business software.
The way you interact with your CRM, your accounting platform, your email, and your calendar in two years' time is unlikely to look like clicking buttons.
It is more likely to look like telling an agent what you want and watching it click the buttons on your behalf.
Whoever owns the agent that does that clicking is in a position of considerable strategic importance over your business.
The second is that the choice between an open-source agent (like OpenClaw, running on your own hardware) and a closed, subscription-based one (like the agents Anthropic, OpenAI, Meta, and Google are now building) is the modern equivalent of the cloud-versus-on-premise decision Australian businesses faced 10 years ago.
Made well, the choice is invisible. Made poorly, it costs significantly more than expected and is very hard to reverse.
The third is that, for a brief and likely-closing window, the economics genuinely favour the open-source path for many small and medium businesses.
The AI models that run on a $1,500 Mac Mini are now good enough for the majority of office work.
The hardware itself qualifies for the current $20,000 instant asset write-off. The supply of that hardware is tight and getting tighter. And the clones being built by Meta and Google are, for now, still in internal testing.
What we'll cover next
This piece has set up the players, the vocabulary, and the stakes.
In our next piece in this thread, we will compare the six AI agents currently competing for Australian businesses' attention, what each costs in Australian dollars once GST and currency markup are included, and where the lock-in risks sit with each.
After that, we will turn to opinion: what we think Australian SMEs should actually do about this, and why the next four weeks may matter more than the next four years.
For now, the take-home is simple. The shift from chatbots to agents is real, the timing pressure is real, and the decisions you make in the next quarter are likely to compound.
If a friend or colleague has been confidently throwing around terms like "OpenClaw", "agent harness", or "orchestration layer" and you have been nodding politely, you now know what they mean.
That is a useful place to start.
Luke Vaughan is the publisher of Business News Australia and an Australian entrepreneur recognised by this masthead in its Top 100 Young Entrepreneurs before acquiring it. He has founded and oversees the operation of several Australian businesses.
A footnote, in the spirit of full disclosure: this article was drafted with the assistance of Claude, the Anthropic AI product mentioned above.
)
