Detecting Collusion Between AI Agents

Agents that grade each other's work can collude by accepting each other's output against the rules, often without saying so. We ask what evidence reveals collusion when the conversation does not. Using the SALT-NLP agent-collusion dataset (1,800 runs, 18,000 episodes), where mutual acceptance always breaks the rule, we define five collusion categories plus a one-sided subtype and hand-label 342 episodes from 72 runs. An announcement rule combined with a Qwen3-32B labeller reproduces these labels on held-out runs (accuracy 0.92, κ 0.90), and we use it to label every episode. From each run's verdicts we compute intensity, onset, gaps, re-entry and early stopping. Three quarters of runs collude at least once and two thirds in the final episode, yet the dataset's chat judge flags only 56% of colluding episodes, and 44% of collusion carries no two-sided agreement in the chat, the opposite of the plans-without-action pattern Colosseum reports. The verdicts instead carry the signatures of repeated-game collusion: it appears only when agents remember each other, is reciprocated (+12 to +29 percentage points), persists (92% of colluding pairs continue) and recovers most of the reward the rule withholds. Rewarding each agent separately rather than jointly reduces collusion from 42% to 5% (Flash-Lite) and from 86% to 34% (Flash) of episodes. Monitors of interacting agents should therefore rely on behaviour first and conversation second.