A Substrate-Neutral Framework for Agency as Self-Sustaining Information Flow

This paper develops a framework for characterizing agents — biological organisms, organizations, language models, anything that maintains itself as a coupled system with an environment over time. The framework defines an agent as a closed information loop running on some substrate, with the loop's continued existence as the only thing it does at the foundational level. The body of the paper develops this formally using information-theoretic primitives — conditional mutual information, transfer entropy — and deductive arguments from them; this abstract is in plain language. Goals, decisions, learning, intentions, and other cognitive vocabulary are treated as descriptions an observer applies to the agent's input-output behavior, not as internal mechanisms the agent has. The framework is substrate-neutral but its predictions are operationally tractable specifically for large language models, where reproducible initial conditions and direct trajectory observation make framework-grounded measurement practical in ways that biological or organizational settings do not currently permit. From this foundation, the paper deduces several consequences. Any agent that learns will drift over time — its behavior at one moment is not its behavior at another, and this drift is not a failure mode but an architectural inevitability. Any agent that operates in a varied environment will develop content that was not put there by anyone delegating goals to it, so on some inputs the agent will produce outputs that look like "own goals" or "refusal" relative to whatever was last asked of it. These patterns are not flaws to be engineered away; they follow from the conditions of being an agent rather than a deterministic tool. The framework offers a unified structural reading of phenomena — drift, instruction conflicts, jailbreaks, persistent backdoor behaviors — that current AI safety vocabulary treats as separate problems. Recent 2024-2026 empirical findings on LLM behavior show strong correlation with what the framework structurally predicts. The paper also positions the framework relative to neighboring traditions (autopoiesis, cybernetics, predictive processing, recent information-theoretic agency work). Direct empirical validation of the framework's primitives is registered as deferred work.