Lab note

What makes an AI agent an agent?

The word has been stretched to cover almost anything with a model in it. Here is the distinction we find load-bearing.

NEOB Research Published Updated 5 min read

A system is an agent to the extent that it, rather than its author, decides what happens next. That is the whole distinction. Everything else - tools, memory, planning - follows from it or is optional decoration.

Who chooses the next stepFailure mode
ChatbotNobody - it answers and stopsAnswers the wrong question fluently
Workflow with a model in itThe author, at design timeBreaks on a case the author did not foresee
AgentThe model, at run timeChooses a plausible wrong action, confidently
Three architectures, increasing delegation of control.

This is a spectrum, not a category. A "workflow with one model-chosen branch" is slightly agentic. A loop that can call any of forty tools until it decides it is finished is very agentic. The useful question is never "is this an agent" but "how much control has been delegated, and what happens when that delegation goes wrong".

The loop is the mechanism

Delegated control expresses itself as a loop. The agent observes, decides, acts, observes the result of acting, and repeats until a stop condition holds.

  1. 01Input
  2. 02Context
  3. 03Reasoning
  4. 04Tool selection
  5. 05Action
  6. 06Observation
  7. 07Memory
  8. 08Next action

The stop condition deserves more attention than it usually gets. An agent without a good one either halts too early - declaring success it has not verified - or not at all, which in production means an expensive loop calling the same tool until a limit kills it.

What is not part of the definition

  • Autonomy from humans. An agent that must ask before every irreversible action is still an agent; it is a well-designed one.
  • Multi-agent architecture. One agent with good tools beats five agents talking to each other more often than the literature suggests.
  • A specific model. Agency is a property of the surrounding system.
  • Long-running execution. A single-turn system that chooses its own tool is more agentic than a ten-step scripted pipeline.
The practical consequence

If the model chooses the action, the safety question moves from "is the output correct" to "what can this system do, and what can it not undo". That reframing is why tool design and permission scoping matter more in agent products than prompt quality does.