How an AI Economy Behaves: A Multi-Agent Simulation in Which AI Companies Learn to Compete by Reinforcement Learning

Abstract

We present an open, reproducible simulation of a stylised AI industry over ten years and let AI product companies learn to compete in it by reinforcement learning (RL). The simulated economy links a chip fabricator, chip designers, data centres, frontier and open-weight model providers, 5–12 competing AI product companies and 300 enterprise customers in three sectors. Companies decide which use cases to pursue, which model to build on, how much to spend, and when to pilot, launch, scale or stop, facing uncertain technology, hidden customer value, compute shortages and rivals. We fixed the training changes, test markets and success bar in a protocol written before the v2 experiments, trained five agents on a laptop in 13–30 minutes each, and tested each on never-used markets. In 100 new markets the agents created $19.4M of value per company against $14.3M for the strongest hand-written rule; four of five were clearly better and none worse, narrowly missing our pre-set bar of all five. They also led under faster or slower technology progress and a wave of entrants; with harder use cases the lead was not statistically clear, and in a severe compute crunch there was no statistically detectable difference from the rule, while the agents went bankrupt more often. Comparison runs found no detectable benefit from training on varied markets, and a single run showed that learning from scratch, without first imitating a rule, can work. Scores on the practice markets used to pick each agent overstated test performance by about $2.8M. In the simulated market, adoption follows S-curves, the regulated sector is held back by reliability, and supply shocks nearly quadruple compute prices. Written for readers new to RL, the paper explains each idea where it first matters, includes a glossary, and states real-world lessons as testable hypotheses. Code, trained agents and a live viewer are open source.

Keywords