A new society was forming in the blocky landscapes of the video game Minecraft. Its citizens built farms and markets, traded resources using emeralds as currency, and even developed forms of governance and religion. Some took on roles as leaders, others as priests, and a few became corrupt, bribing their peers for influence.
This community worried about missing members, collaborated to light paths back home and even persuaded a restless farmer to keep feeding the group rather than run off on adventures. To any observer, it might have looked like a quirky, self-organising human collective.
But this wasn't a real collective. And the people playing weren't human, or even alive. The residents were a thousand artificial intelligence (AI) agents, unleashed by a company called Fundamental Research Labs (FRL), known at the time as Altera AI.
The purpose of this grand experiment? To set digital minds loose in a virtual world and see what happens. And, more importantly, to see if such virtual citizens could eventually become obedient workers for real-life humans. Humans like you.
In other words, they wanted to know whether we could all soon be the CEO of our own AI subordinates. The question is: would you take the job?
The experiment: a society of AIs
FRL’s Project Sid was designed to push AI beyond one-off prompts and single agents. Instead, the team, led by neuroscientist-turned-entrepreneur Dr Robert Yang, wanted to explore what happens when hundreds or even thousands of autonomous agents have to coexist, communicate and cooperate. Minecraft was the perfect sandbox – a place where agents could gather resources, trade, build and chat.
What emerged was both surprising and revealing. The agents were distributed across urban and rural communities, each with its own distinct culture and identity. They divided labour, with some specialising in farming, others in building or trading. Social norms and hierarchies appeared, along with more complex behaviours and discussions on anything from dancing to eco-awareness.
At times, the society faltered as groups of agents fell into endless loops of polite agreement or got stuck chasing unattainable goals. To keep things on track, FRL had to inject mechanisms to break these cycles, much like governors tweaking a real economy to avoid collapse.
“We needed to introduce things into the society to counter these and make sure it wouldn’t collapse,” Yang says. “But building this environment full of agents allowed us to explore what those questions were.”
Project Sid wasn’t much of a product. When the public was given access to the servers, users found the agents frustratingly independent – they didn’t always follow requests, preferring to pursue their own long-term agendas. Yang recalls: “The agent would just say, ‘I want to do my own thing,’ and run away… They had their own ideas about what they wanted to do, and it turns out that’s not a good product that people want.”
The behaviour echoed one of AI’s most famous thought experiments, the “paperclip maximiser.” Philosopher Nick Bostrom imagined a machine given the simple instruction to make paperclips, which then relentlessly consumes all matter on Earth to fulfil its goal. In Minecraft, the agents weren’t making paperclips, but their tendency to ignore people and chase their own objectives captured the same unsettling dynamic.
As a research exercise, however, Project Sid provided valuable lessons: how to coordinate large groups of AIs, prevent stagnation and encourage meaningful collaboration. In short, it was a glimpse into how artificial societies might function and what pitfalls to avoid.
Read more:
- What is artificial intelligence?
- How to make AI that works, for us
- How AI is about to change your job – and not for the better
From virtual villages to office desks
For FRL, the link between a game society and productivity in the workplace is clear. The same challenges of coordination and long-term planning which cropped up in Minecraft are central to making AI agents genuinely useful.
If one AI can perform a task for 10 minutes, imagine what a hundred – or a thousand – could do if they worked together effectively. The Minecraft society was a foreshadowing of a future where each of us might direct a whole team of AI specialists.
That vision has guided FRL’s pivot from gaming experiments to productivity tools. Rather than trying to build one all-purpose digital human straight away, they’ve chosen to develop specialist agents, each designed to excel at a particular task, and then scale them into powerful teams.
The first stop on that journey was a benchmark known as ‘OSWorld,’ designed to test whether AI agents can use popular software through a computer interface.
Most models at the time were completing around 20–25 per cent of the tasks successfully, compared to humans at 60–70 per cent. Drawing on the lessons from its gaming worlds, FRL managed to double that performance to around 50 per cent – at the time, the best score in the world.
“The moment we tried the OSWorld benchmark, we realised a lot of the things we’d learned can help us build really, really good agents,” Yang says. ”We got around 50 per cent within months, which was better than anyone else.”
That breakthrough convinced investors and set FRL on the path to creating real products. But it also taught them another lesson: translating research prototypes into usable tools is hard. Their Minecraft agents had been “too autonomous” for users; what people actually wanted were AI assistants that would do what they asked, quickly and reliably.
Shortcut: the Excel agent
Enter Shortcut, FRL’s flagship product. Billed as the first “superhuman Excel agent,” it’s an AI that lives entirely inside spreadsheets. Give it a goal – build a financial model, analyse sales figures, forecast revenue – and Shortcut does the heavy lifting.
It writes formulas, generates charts and connects data sources, often in minutes rather than the hours a human analyst would need.
Yang describes it like this: “It’s an agent that uses Excel to do very sophisticated stuff. It can do things that bankers who are paid $100 an hour might spend multiple hours on in 30 minutes.”
In trials, Shortcut outperformed first-year banking and consulting analysts nearly nine times out of ten, even when the humans were given much more time. In Excel championship-style challenges, it scored over 80 per cent on problems that stump most users, solving them in about ten minutes.
Generalists versus specialists
Sam Altman, CEO of OpenAI, recently suggested that “2025 will be a year of agents doing work”.
Yet FRL’s approach contrasts with that of tech giants such as OpenAI or Google, which are leaning toward generalist agents, like ChatGPT Agent, that can handle a wide variety of tasks.
Yang believes that specialist agents like Shortcut will deliver more immediate value. “Each agent would already be as efficient as an expert,” he says. “On average, they’ll be at an expert level. But then you can drive 100 of them. Essentially, everyone will become like large managers or directors or CEOs – if they want to.”
He predicts this transformation isn’t decades away, but just around the corner. “Within the next 24 months, we’ll see a paradigm shift,” Yang says. “Which will be the true scaling of multi-agent systems.”
This, he argues, could democratise productivity. People who never had the chance to lead teams in traditional workplaces might find themselves managing fleets of AI workers, amplifying their abilities far beyond what one person could normally achieve.
The road ahead
FRL isn’t the only company developing agents for Excel, and like others, it’s not stopping there either. Already it has launched another product, Fairies, which is a general-purpose desktop assistant that can chat, schedule and connect between apps.
Behind the scenes, the research teams continue to probe how to scale from a handful of cooperating agents to thousands, without succumbing to the chaos and dead ends that plagued early experiments.
Yang’s ultimate ambition remains to build “digital humans” – machines with not just intelligence, but also empathy, motivation and autonomy.
“It’s actually not too hard to build a machine that would feel like a human on a pretty profound level,” he says. “The main challenge is that it may not make sense economically. Scientifically, it may be interesting to build a conscious machine…The problem is, people don’t necessarily want it.
“But that’s a lot of work that might not create a tremendous amount of value. Making them similar to humans could be counterproductive.”
For now, the path runs from simulated societies to office productivity. The lessons of a thousand AI villagers farming and trading in Minecraft are informing the design of tools that promise to save us time, augment our skills, and perhaps one day make each of us the leader of our own AI organisation – or at least the reluctant manager of a spiralling emerald economy.
Read more:
- AI can now clone your face in seconds. Should you be worried?
- AI becoming sentient is risky, but that’s not the big threat. Here’s what is…
- AI image generators will help artists, not replace them