Unit 01.01: Scripted workflows that get called agents

A large share of what gets deployed as agentic automation is a script with a model in the middle, and the mislabelling costs real money.

Applying the test to five workflows

For each, ask whether the sequence of steps depends on something computed at run time.

The table below sorts five workflows.

workflow                                           agentic?
read a form, summarise it, file the summary        no -- it is a script
classify a ticket, route to one of four queues     no -- it is a script
look something up, then decide whether to look up more yes
draft an email in house style                      no -- it is a script
keep searching until you find an answer or give up yes

3 of 5 are scripts with a model inside them

The ticket-routing case is the interesting one. It looks agentic - a model makes a decision, and different tickets go different ways - and it is a script. The four destinations and the rule that picks between them both exist before the run starts.

That distinction matters because the two need different engineering. A script needs its classification tested; an agent needs stopping rules, a loop detector, a permission model and a way to reconstruct what it did.

The mistake this prevents

The mistake is building agent machinery around a script, which is expensive and adds failure modes the script did not have. If the sequence is fixed, a plain procedure is cheaper, faster and easier to reason about - and it cannot loop.

Takeaway

Three of five common workflows are scripts. Apply the test before choosing an architecture: a fixed sequence needs none of what the rest of this course builds.

Mark complete