Agents Trapped

Agents get trapped in self-generated signal loops when their coupling channels drown out the external world — and ant mills, social-media cascades, and addiction are all the same trap.

The core intuition is simple: an agent is supposed to track the world. It does this by building a model, evaluating outcomes, and planning actions. But if the signals feeding that model come mostly from the agent's own past outputs — laundered through pheromones, algorithmic feeds, or sensitized neural circuits — the agent ends up regulating itself instead of reality. The paper formalizes this as a reversal of a single inequality: the agent's model should share more algorithmic information with the external world than with its own prior outputs. When that flips, you get a pathological loop.

What's clever is the module-level diagnosis. Each of the three cases breaks at a different place. The ant mill is a modelling engine failure: the ant's world model is too narrow to represent the possibility that the trail it's following is one it helped create. The social-media cascade is an objective function failure: engagement and emotional salience get mistaken for truth, and the planning engine stops asking "would I believe this if I hadn't seen it six times?" Addiction is primarily a planning engine failure: a cached, cue-driven policy decouples from deliberation, while the objective function's "wanting" signal sensitizes upward even as "liking" stays flat or falls. The modelling engine then narrows around cue-rich states, closing the loop from the inside. None of these are bad designs in their original context — they all become pathological only when the input channel gets colonized by self-generated structure.

The paper then scales this up to collectives. A healthy multi-agent system — a "GAR equilibrium" — is one where every agent's model is still dominated by the external world, not by what its neighbors are outputting. The structural condition that keeps this true is excitation/inhibition (E/I) balance on the network's coupling graph: amplifying connections must be counterweighted by attenuating ones, so self-generated signals decay rather than compound. Pheromone evaporation, accuracy friction on platforms, and prefrontal inhibitory control are all substrate-specific instances of the same negative-feedback requirement. Lose that balance — through sensitization, algorithmic amplification, or weakened inhibitory control — and the collective becomes a good regulator of itself rather than of the world.

Two sections of the paper are explicitly flagged as unfinished drafts ("to be drafted in the next pass"), so the design-principles section and the corpus-connections section are placeholders rather than developed arguments. The theoretical scaffolding is complete; the applied and connective work is not yet there.