agent definitions — artificial intelligence

overview

ai definitions anchor agents in perception–action loops, planning, and learning. the cluster ranges from classical textbooks to documentation for llm-based agent toolkits released in 2025.

signature traits

• environment interaction: agents sense, reason, and act to maximize performance measures (russell & norvig, 1995).
• autonomy and proactivity: franklin & graesser’s taxonomy, AOSE roadmaps, and MAS texts stress autonomous decision-making.
• tool use & orchestration: modern frameworks (langchain, autogen, anthropic, openai) highlight iterative tool calls, state tracking, and guardrails.

illustrative definitions

• 1995 — russell & norvig, artificial intelligence: a modern approach: an agent perceives via sensors and acts via actuators to maximize a performance measure.
• 1996 — franklin & graesser, “is it an agent, or just a program?”: autonomous agents pursue their own agenda over time within an environment.
• 2025 — openai, anthropic, langchain, autogen docs: define agents as llm-driven systems that plan, call tools, coordinate, and deliver results autonomously.

relation to other dimensions

• autonomy spectrum: definitions trend high—autonomy is assumed and measured by the agent’s capacity to handle tasks end-to-end.
• entity frames: machine-centered, though some frameworks integrate human oversight loops (hybrid).
• goal dynamics: span acceptance (optimize provided goals) through negotiation in multi-agent orchestration.
• persistence & embodiment: mostly digital persistence; robotics contexts add physical embodiment when actuators are real-world devices.

open questions

• how do llm-based agents change evaluation metrics compared with classical reactive or deliberative agents?
• what governance mechanisms ensure autonomous tool-calling agents behave safely in open environments?
• can we standardize terminology across vendors to avoid fragmenting the “agent” concept further?