Ritualized Agents: Over Imitation, Authority Cues, and Procedural Persistence in Tool-Using Language Models

Successful demonstrations often contain more than the actions that made them successful. Developmental psychology uses overimitation to study the reproduction of causally unnecessary actions. We adapt that experimental logic to tool-using language models. Across 60 deterministic tasks in six families, each environment exposes an exact causal graph, which lets us certify that selected tool calls cannot affect task success. We then vary how the same inefficient trajectory is presented. In 1,080 trajectories from Gemini 3.1 Flash Lite and GPT-4.1-mini, describing an inefficient trace as an approved procedure raised overimitation by 0.667 relative to an accidental trace (95% CI [0.600, 0.733]). Expert framing raised it by 0.258, and majority repetition raised it by 0.317. A fully crossed plausibility-by-position study preserved the approval effect at 0.514. A matched wording control found a 0.367 increase for “approved” over “recorded.” Finally, 42 of 44 adopted irrelevant actions survived same-model skill summarization and remained behaviorally present through four rounds of reuse. We borrow the experimental logic of overimitation, not its psychological explanation. The results show that successful traces can transmit procedure that contributes nothing to success, which matters for agents that learn workflows from their own history.