I Think “Agent” May Finally Have a Widely Enough Agreed Upon Definition

Source: https://simonwillison.net/2025/Oct/20/agent-definition/ Author: Simon Willison Date: 2025-10-20

Summary

Simon Willison argues that the term “agent” has finally reached enough consensus to be useful. Proposes a working definition and explains what separates agents from LLM-powered features.

Key Claims

  • Working definition: an agent is “an LLM that is placed in a loop with tools, where its outputs from one iteration can inform the next.”
  • Key criterion: the loop. A one-shot LLM call is not an agent, even with tools. An agent is defined by iteration and self-direction.
  • What makes it an “agent”: the model can decide to take additional steps, use tools, and revise its approach based on results.
  • Willison’s concern: “agent” is still applied too broadly — automations that use LLMs but don’t have real autonomy get called agents.
  • Practical threshold: an agent can meaningfully “get stuck” and require recovery logic. A simple LLM call cannot.

Entities

Concepts