Skip to content
NTPU AI Learn

AI 104 · Module 1

From Task to Agentic Workflow

What is the difference between a single prompt, a fixed workflow and an agent?

Start Section 1
  • Duration30 min
  • LevelCore
  • PrerequisiteAI 103 recommended (Memory, Knowledge & Skills)
0 / 4 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor

Info

After you join, your completed lesson numbers are sent to your instructor together with your learner code, so they can see your progress. Use the codes your instructor gave you, and do not enter your name, company secrets or other personal data. Progress is kept with the rest of that session's data.

Learning objectives

  • Distinguish a single Prompt, a Chain, a Workflow, an Agentic Workflow and an Agent
  • Explain why repeated work should not always be handled as isolated prompts
  • Use "who decides the next step?" to judge whether a Workflow is enough or an agentic approach may help
  • Choose one real work task, described with simulated data, to use throughout the course

Problem and need

8 min

Re-issuing the same instructions every week

In AI 103 you learned what an agent needs to remember, know and be able to do. Even with those capabilities, one question is still open: how should the agent carry out the work step by step? The core question of this course is: how does an agent complete a task step by step?

Work scenario (simulated)

An administrative officer summarizes course feedback every Friday. Each time she has to instruct the AI in turn: paste the feedback, explain the categories, ask for a summary, ask for a list of issues that need follow-up, and finally ask it to check for anything missing. This week she forgot the "check for omissions" step, and the report missed an important complaint.

  • The steps and their order live in memory, so they get skipped or vary from week to week
  • A colleague cannot take over, because the method exists only in one person's head
  • There are no fixed checkpoints, so when something goes wrong it is hard to tell which step failed
  • Every run starts with the same explanation, which costs more time than it seems

Warning

The opposite mistake is just as common: handing everything to a single autonomous agent. More autonomy means less predictable behavior, higher cost and harder review. The goal is to choose a design that is just enough for the work, not the most impressive one.

Key takeaway: Repeated work deserves a repeatable process, but choose just enough autonomy rather than as much as possible.

Core concepts

8 min

The spectrum from Prompt to Agent

From a single prompt to an agent is a continuous spectrum. The further right you go, the less the next step is written in advance and the more the AI chooses it based on the goal, context, state or result.

The spectrum from Prompt to Agent
  1. Prompt: one instruction, one response
  2. Chain: several prompts linked in a fixed order, each output feeding the next input
  3. Workflow: predefined steps, branches and checkpoints
  4. Agentic Workflow: a fixed overall flow in which some steps are judged by the AI within limited options
  5. Agent: the next step may be chosen by the AI based on goal, context, state or result

Text equivalent: the five stages run from left to right in a line. A Prompt is the simplest form; a Chain links several prompts; a Workflow adds branches and checkpoints; an Agentic Workflow leaves a few places in a fixed flow for the AI to judge; and in an Agent even the next step is chosen by the AI.

Comparing the five stages (simulated examples)
StageWho decides the next step?Example
PromptNot applicable, there is only one step"Shorten this text to 100 words."
ChainWritten in advance, fixed orderTranscript → summary → translation, three prompts in sequence
WorkflowDefined in advance, may include conditional branchesFeedback → classify → summarize → draft report → human review
Agentic WorkflowMostly defined in advance; a few steps are judged by the AI within optionsThe flow is fixed, but "does this comment need follow-up?" is judged by the AI, and unclear cases go to a person
AgentThe AI chooses based on the goal and the current situationThe goal is "find out why satisfaction dropped this term", and the AI decides which data to look at
Check yourself

A process that runs every Friday has five fixed steps in the same order every time, each executed by AI, with a person reviewing at the end. Which is it closest to?

A process that runs every Friday has five fixed steps in the same order every time, each executed by AI, with a person reviewing at the end. Which is it closest to?

Key takeaway: Prompt to Agent is a continuous spectrum; the test is who decides the next step.

8 min

Workflow vs. Agent: who decides the next step?

Workflow

  • The next step is mostly predefined
  • The same input follows the same path
  • Easy to predict, test and review
  • Suits repeated work with clear rules that needs consistency

Agent

  • The next step may be chosen based on goal, context, state or result
  • The same goal may take different paths
  • More flexible, but harder to predict and check
  • Suits open-ended work where every step cannot be written in advance
Deterministic vs. agentic in five dimensions
DimensionDeterministic (fixed)Agentic (judged by AI)
RepeatabilityNearly the same every timeMay differ each time
Decision autonomyDecided in advance by the designerDecided by the AI within allowed limits
Execution stateOnly needs to know which step is nextNeeds to track the current situation and what has been done
Human oversightReview at fixed checkpointsSet boundaries, and have a person step in at key points
Cost of a mistakeEasier to notice and traceHarder to trace, needs more safeguards

Info

A practical starting point: if you can define the next step in advance, use a Workflow. Only where you cannot, and the AI can choose safely among limited options, add an agentic judgment to that step. Deeper agent design, such as Planning, belongs to later courses.
Check yourself

A support mailbox sends each message by category (billing, technical, general) to one of three predefined handling flows, and every flow has fixed steps. Is this an Agent?

A support mailbox sends each message by category (billing, technical, general) to one of three predefined handling flows, and every flow has fixed steps. Is this an Agent?

Key takeaway: In a Workflow the next step is mostly predefined; in an Agent it may be chosen by the AI from the goal and situation.

Practice and apply

6 min

Practice: Prompt, Workflow or Agent?

Below are 8 simulated scenarios. Choose one for each: Prompt, Workflow, Agentic Workflow or Agent, then discuss with a partner or the instructor. Some cases can reasonably be read two ways; what matters is that you can explain your reasoning. Please do not enter real company secrets, client personal data or any unauthorized sensitive information in this exercise.

8 scenarios (decide for yourself)
CaseScenarioThink about
1Ask an AI to shorten an announcement to 100 wordsIs it only one step?
2Every Friday: read feedback → classify → summarize → draft the report, always in the same orderIs the next step defined in advance?
3Transcript → summary → translation → formatting, each output feeding the next stepIs this the simplest fixed flow?
4A purchase request follows a fixed flow; only the step "is the attachment enough to decide?" is judged by an AI, choosing between asking for more or sending on for reviewMostly fixed, with a few AI-judged steps
5Goal: "investigate three competing options and recommend one"; the AI decides what to search, what to read and when it has enoughWho decides the next step? Where are the limits?
6New-student questions are routed to one of three prewritten reply templates by question typeAre the branch conditions defined in advance?
7A literature review has five fixed steps; "is this paper relevant?" is judged by an AI against criteria, and unclear cases go to a personA fixed flow with a place for AI judgment
8Ask an AI for five ideas for an event nameDoes it need several steps or checkpoints?
  • Prompt: one instruction, one response
  • Workflow: the next step is defined in advance (including a Chain and conditional branches)
  • Agentic Workflow: mostly fixed flow, with some steps judged by the AI within limited options
  • Agent: the next step may be chosen by the AI based on goal and situation
  1. Apply: choose one real task for the whole course
    Pick a task you have actually done, that recurs, and that has at least three steps. Every later module practices on it, and it ends up in your Agentic AI Workflow Canvas.
  2. Describe it with simulated data
    Write only the steps and nature of the work, with no real names, client personal data or company secrets. If the task involves sensitive content, use a simulated version instead.
  3. Write the goal in one sentence
    For example: "Produce a weekly course-feedback report that a manager can review."
Check yourself

You need to pick one task for the whole course. Which is the best fit?

You need to pick one task for the whole course. Which is the best fit?

Key takeaway: Human-AI-Human: a person defines the task and its limits, the AI assists or executes along the flow, and a person reviews and stays responsible for the result.

Save as PDF

What next?