agent definitions — cognitive science

published: October 27, 2025

overview

cognitive-science accounts treat agents as building blocks of cognition: modular processes, intentional planners, or distributed mechanisms that collectively form intelligent behavior. agency becomes a lens on mental architecture, not just overt action.

signature traits

  • modular societies: minsky’s society of mind frames agents as simple processes whose interactions yield complex cognition.
  • plan-based control: bratman and related theorists emphasize partial plans and practical reasoning structures.
  • representational focus: agency is tied to information processing, symbolic manipulation, or distributed representations inside the mind.

illustrative definitions

  • 1986 — marvin minsky, society of mind: intelligence emerges from a “society” of small agents; agency is micro-level processing.
  • 1987 — michael bratman, intention, plans, and practical reason: agency consists of stable, future-directed intentions organizing cognition and action.
  • later cognitive architectures: bdi (belief-desire-intention) models bridge cognitive science with agent-oriented programming, keeping mentalistic vocabulary.

relation to other dimensions

  • autonomy spectrum: autonomy is internal—agents manage their own plan hierarchies, emphasizing self-organization over external delegation.
  • entity frames: begins human-centered but readily generalizes to machine agents by abstracting minds into process societies.
  • goal dynamics: supports adaptation and limited negotiation among internal sub-agents coordinating plans.
  • persistence & embodiment: focuses on persistent mental structures rather than physical embodiment; embodiment is optional if representation suffices.

open questions

  • how do modern neural architectures map onto the “society of mind” metaphor—are attention heads agents?
  • can we operationalize partial plans and BDI constructs for contemporary llm-based agents without hand-crafted symbols?
  • what cognitive constraints should we import when designing oversight interfaces for multi-agent llm systems?
on this page