Unit 01.00: The one property that makes something an agent

"Agent" is used for almost everything an LLM touches. There is exactly one property that distinguishes an agent from a script, and it is not the one most definitions lead with.

Does anything decide the sequence while the run is happening?

A script runs a fixed sequence. An agent decides its next step from something computed during the run. That is the whole distinction.

The comparison below sets out the same three steps written both ways.

SCRIPTED WORKFLOW                     AGENTIC WORKFLOW
step 1  check the invoice             step 1  check the invoice
step 2  approve if under 500          step 2  ...decided after step 1 returns
step 3  post the payment              step 3  ...decided after step 2 returns

the sequence is in the design         the sequence exists only after the run

  Same three steps. In the first, you can read the order off the page. In the
  second you cannot, because the order is chosen while the run is happening.

The scripted version's order is in the source file. The agentic version's order is in trail, which only exists after the run - you cannot read the sequence from the code because the code does not contain it.

Notice what appears in neither column: a model. A scripted workflow with four model calls is still scripted, and an agent whose next-step rule is a simple lookup is still an agent. The property is about control flow, not about intelligence.

The mistake this prevents

The mistake is defining agents by the technology inside them - model calls, tools, a framework. Those tell you what the system is built from. The run-time-sequence property tells you what you have to build around it: stopping rules, permissions, a way to see what it did.

Takeaway

An agent is a system whose next step is decided during the run. Everything else in this course - stopping rules, permissions, gates - exists because of that one property.

Mark complete