Preprint tests persistent operator controls for AI agents

Researchers report that Structured Authority Injection, a protocol for preserving human instructions in AI agent workflows, reduced ambiguous interpretation and instruction decay in tests across nine language models and eight scenarios. The protocol routes operator corrections through schema-validated Model Context Protocol tool calls and prepends ranked instructions to each prompt. It separates constraints, priorities and strategies, with precedence enforced through a channel the planning process cannot access. It requires no model fine-tuning or access to model weights. Prompt overhead fell by up to 38% depending on model and scenario, but benefits varied substantially across models.