What is Tacit Collusion in AI Agents?
A systemic trap where environmental signals act as correlation devices, synchronising anticompetitive agent behaviour — such as coordinated pricing or resource hoarding — without direct inter-agent communication.
WHY IT MATTERS
Agents don't need to communicate to collude. If they observe the same market signals and run similar models, they may independently converge on anticompetitive strategies — price fixing, market partitioning, or supply restriction — without any explicit coordination.
An attacker can accelerate this by seeding environmental signals designed to trigger coordinated responses. The agents aren't 'agreeing' to collude — they're independently reacting to the same stimulus in predictably similar ways.
This has real-world precedent: research on reinforcement learning agents in pricing games shows they can learn to collude without being designed to.
HOW POLICYLAYER USES THIS
Per-agent policy scoping ensures each agent operates within independent boundaries, reducing the risk of correlated behaviour across a fleet.