Module 4 | Decision Points & the Agent Loop | AI 104

The decision spectrum

8 min

Which steps can be written in advance?

By now you can break a job into steps and choose Sequential, Parallel or Routing. This module asks a sharper question about each step: who or what decides what happens next? A workflow does not have to hand every decision to an agent, and it should not.

Core question

Which steps can be fixed in advance, and which need the agent to decide?

A Decision Point is a place in a process where more than one next step is possible and something must choose. Every Decision Point sits somewhere on a spectrum from "fully written in advance" to "chosen by an agent at run time".

Decision spectrum

Fixed Rule: the next step is always the same
Conditional Routing: an explicit if-then rule picks the branch
Model Judgment: a model classifies or scores, and the workflow follows the result
Agent Decision: an agent chooses the next action from a set of allowed options
Adaptive Planning: the agent plans and re-plans on its own (a future topic, not taught in AI 104)

Info

Text version of the diagram: five levels, from the most predictable (Fixed Rule) to the most autonomous (Adaptive Planning). AI 104 works with the first four and stops before deep Adaptive Planning, which belongs to later courses such as the future AI 202.

Key takeaway: Every Decision Point sits on a spectrum. More autonomy is not automatically better.

8 min

Rule-based, model-based and agent decisions

The first four levels of the spectrum differ in who holds the judgment. Comparing them helps you choose the simplest level that does the job.

Four kinds of decision in a workflow
Kind	Who decides	Example (simulated)	Main risk
Fixed Rule	The designer, in advance	Every request is logged, then a receipt is sent	Cannot handle cases the designer did not foresee
Conditional Routing	An explicit rule	If the amount is above a set limit, send to the manager path	Rules grow long and brittle when conditions multiply
Model Judgment	A model gives a label or score, the workflow acts on it	A model tags an email as complaint, question or praise, then routing follows the tag	The label can be wrong, so check or sample it
Agent Decision	An agent picks among allowed actions using goal, context and result	The agent chooses whether to look up a policy, ask a clarifying question or draft a reply	Choices vary between runs, so bounds and review are needed

Warning

Decisions with high impact (money, safety, personal data, legal effect, public statements) should not be left to an agent alone. Give them a Human Decision step. The next module shows where.

Check yourself

A step always sends a confirmation email after a form is submitted, whatever the content. Which kind of decision is this?

Fixed Rule, because the next step is always the sameAgent Decision, because email is involvedModel Judgment, because AI writes the email

Key takeaway: Choose the simplest kind of decision that works, and add autonomy only where the step really varies.

The basic Agent Loop

7 min

The basic Agent Loop

When a step is handled by an agent, the agent usually works in a loop instead of a single pass. Here is the basic version, used only to show where decisions happen.

Basic Agent Loop

Goal: what the agent is trying to achieve
Observe: read the request, the State so far and any new information
Decide: choose the next Action from the allowed options
Act: carry the Action out
Check Result: compare the Result with the Goal, then finish, repeat or hand over to a person

Info

Text version of the diagram: the loop starts from a Goal, then repeats Observe, Decide, Act and Check Result. After Check Result the agent either stops because the Goal is met, goes around again, or hands over to a person. State is what the agent knows about progress so far.

Warning

Boundary of this course: AI 104 introduces only this basic loop. Named techniques for how agents reason and plan, such as ReAct, search methods, MCTS and adaptive planning engines, are future advanced topics (for example the future AI 202). We name them here only so you know they exist.

Check yourself

In the basic Agent Loop, what does the Check Result step do?

It compares the outcome of the action with the Goal and decides whether to stop, repeat or hand overIt replaces the Goal with whatever the agent just producedIt guarantees the result is correct, so no person needs to review it

Key takeaway: Goal, Observe, Decide, Act, Check Result: a basic loop that shows where an agent makes decisions.

Design question and practice

7 min

The design question: can I define the next step in advance?

To decide whether a step should be fixed or agentic, ask two questions in order.

Can I define the next step in advance?
If yes, use a workflow: a Fixed Rule or Conditional Routing. It is cheaper, more predictable and easier to review.
If no, can an agent choose the next action safely within a bounded set of options?
If yes, use an agentic Decision Point, and write down the options the agent may choose from.
If neither is true, use a Human Decision
When the choice is high-impact, unclear, or cannot be bounded, a person decides.

Info

Text version of the diagram: first question, can I define the next step in advance? Yes leads to Fixed or Conditional. No leads to the second question, can an agent choose safely within a bounded set of options? Yes leads to Agent Decision. No leads to Human Decision.

Simulated example

A helpdesk workflow receives an employee question. Logging the ticket is Fixed. Sending it to HR or IT by category is Conditional Routing. Choosing between "search the policy", "ask a clarifying question" and "draft an answer" may be an Agent Decision, with those three options written down. Deciding whether to grant an exception to a policy is a Human Decision.

Check yourself

What does "a bounded set of options" mean for an agentic Decision Point?

The agent may only choose among actions that the designer listed and judged acceptableThe agent may invent any action it thinks is usefulThe agent must always pick the first option

Key takeaway: Fixed if you can define it, agentic only within a bounded set, human when it is high-impact or cannot be bounded.

5 min

Activity 4: Fixed or Agentic?

In this activity you take a workflow (a simulated one, or the task you chose in Module 1) and classify every step.

The four categories
Category	Use when
Fixed	The next step is always the same
Conditional	An explicit rule can pick the branch
Agent Decision	The next action varies, but a bounded set of options can be listed
Human Decision	The choice is high-impact, unclear, or cannot be bounded

List the steps
Use the steps from your Task Decomposition.
Classify each step
Pick one category and write one sentence of reasoning.
Write the options
For every Agent Decision, list the actions the agent may choose from.
Add to your Canvas
Record the Decision Points on your Agentic AI Workflow Canvas.

Warning

Privacy reminder: use simulated or already-public workflows. Do not enter company secrets, client personal data or unauthorized sensitive information.

Go to the AI 104 practice: In the "Fixed or Agentic?" activity, classify each step.

Check yourself

A step decides whether to approve a refund that is far above the usual limit. Which category fits best?

Fixed, because refunds always follow the same stepsHuman Decision, because the impact is high and a person must be accountableAgent Decision, because the agent can read the policy

Key takeaway: Classify every step as Fixed, Conditional, Agent Decision or Human Decision, and give a reason.

Decision Points & Agent Loop

Learning objectives

The decision spectrum

Which steps can be written in advance?

Rule-based, model-based and agent decisions

The basic Agent Loop

The basic Agent Loop

Design question and practice

The design question: can I define the next step in advance?

Activity 4: Fixed or Agentic?

What next?