In 2025, Fundamental Research Labs (FRL), known at the time as Altera AI, set 1,000 AI agents loose in a world to see what happened. The goal was to see if they could coordinate large groups of AIs, prevent stagnation, and encourage meaningful collaboration in a way that would make them genuinely useful if they were agents in the real world working for real humans.

What emerged was both surprising and revealing. These early AI agents were distributed across urban and rural communities, each with its own distinct culture and identity. They divided labour in farming, others in building or trading.

At times, the society faltered as groups of agents fell into endless loops of polite agreement or got stuck chasing unattainable goals (common in early AI). Users also found the agents frustratingly independent and often ran off to do their own thing. At first, they were only completing 20-25% of their tasks, which they got up to 50% – but still far short of the 60-70% rate of humans.

What they learned was fully autonomous agents were actually not very good. Instead, they started focusing agents to do more specific tasks. They then developed Shortcut, an excel agent able to take a business goal and then build complex financial models, analyze sales figures, forecast revenue, etc. It was tested and could do first-year banking tasks 9 out of 10 times – proving specialization was far better than generalization.

It’s a study of early AI worth reading.