CogEEGAgent Automates Cognitive EEG Analysis with Bounded Autonomy.

Analyzing electroencephalography (EEG) data in cognitive studies is a complex task requiring specialized expertise and numerous methodological decisions. The CogEEGAgent proposes an AI agent solution, built on MNE-Python, that translates natural language questions into specific analysis choices, offering a flexible automation interface. A key innovation is its "scientific harness" which strictly separates the AI's language understanding from the scientific validation process. This separation ensures that while the AI interprets user intent and suggests analyses, deterministic components validate the proposed methods, control access to confirmatory data, and authorize the release of evidence-bound results. The system demonstrated superior accuracy in mapping language to registered analyses compared to deterministic routers, and its preflight mechanism ensures abstention when required. Policy stress testing further confirmed that its held-out confirmation mechanism effectively curbs false positives from adaptive search, establishing a framework for auditable and bounded autonomous cognitive-EEG workflows.