What Is an Agent?

An agent is a system that takes in information, selects actions, and can affect what happens next. In reinforcement learning, this matters because the system participates in a loop that changes future observations.

Agenticness is a spectrum. A thermostat has a narrow form of agenticness: it observes temperature and turns heating or cooling on. A robot navigating a room has richer agenticness because it must perceive, choose, move, adapt, and pursue goals over time. A language model using tools can become more agentic when it chooses actions that affect external systems.

What is an agent?

Tune the agenticness spectrum

Agenticness can be treated as a spectrum: environmental interaction, autonomy, goal-directed behavior, learning/adaptation, and temporal coherence.

Environmental interaction70%

Autonomy45%

Goal-directed behavior65%

Learning and adaptation40%

Temporal coherence35%

Agenticness score51%

Thermostat

Vacuum robot

Game-playing RL agent

Research assistant agent

This definition makes room for simple threshold agents, RL agents, and tool-using AI systems without pretending they all have the same level of agenticness.

ℹ

The Useful Question

Ask what the system can affect, what it can observe, what it can learn from, and what it is pursuing over time.

Checkpoint

Which change would usually make a system more agentic?

←PreviousAgent-Environment LoopFrom AlphaZero to RLHF Next→States, Actions, Rewards, and PoliciesFrom AlphaZero to RLHF