Adding Decision-Making to Agents

5.3 Adding Decision-Making to Agents
LOs: [5c]
Once an agent can complete a basic task, the next step is to help it decide what to do when conditions change. This is important because many workplace tasks are not one straight line. An agent behavior is the pattern that guides how an agent responds, chooses between options, and continues toward a goal.
As you saw in Unit 3 when you studied conditional logic, workflows can follow different paths based on rules. This subunit builds on that idea by showing how an agent can use similar branching, but with more flexibility in judgment. Instead of only checking whether one condition is true, the agent can weigh information and choose among several actions.
The source explains that decision-making agents can support multi-step task execution. This means the agent does not stop after the first response. It can review input, choose an action, check the result, and move to the next step. That structure is useful when the task requires more than one decision before completion.
One real-world instance in the source is a travel-planning assistant that helps a user decide among options based on needs and limits. The agent can compare choices, ask for missing details, and produce a plan that fits the stated goal. That example shows how decision-making supports a task that is not fully known at the start.
Building on the branching workflows from Unit 3 and the responsive AI work from Unit 4, you can now ask how an agent should behave when it needs to choose, revise, and continue. As you read, consider how decision-making changes an agent from a simple responder into a tool for multi-step work.
Attribution
Title | Author | Source | License
PRDV000 Unit Readings | Saylor Academy | https://learn.saylor.org/ | CC BY 4.0

OUTPUT LICENSE: CC BY 4.0

5.3: Adding Decision-Making to Agents
LOs: [5c]
When you add decision-making to an AI agent, you give it more than a script. You give it a way to evaluate input, choose between options, and continue through a multi-step task until it reaches a result or a stopping point.
This section builds on 5.2, where you created a simple agent with a narrow goal. Here, the focus shifts to behavior design: what the agent should do first, what it should do next, and when it should stop and ask for human review.
A useful agent often follows a pattern such as read, sort, decide, act, and report. For example, a research-support agent might receive a topic, decide whether the request is factual, creative, or unclear, gather the right type of response, and then return a short summary for review. This kind of behavior helps the agent support human decision-making instead of replacing it.
The more steps you add, the more important it becomes to define rules for stopping. An agent should know when it has enough information and when it should pause. Without that boundary, it may loop, guess, or continue with weak input.
Job-ready Tip: Design the stop rule first
When you design agent behavior, decide in advance what should make the agent stop and hand work back to a person. A clear stop rule protects quality and keeps the workflow manageable.
As you read this section, consider how an agent can support a decision without making the final choice itself, and explain what stopping point would make that support safe and useful.
SOURCE: [5.3 Adding Decision-Making to Agents]
ATTRIBUTION:
AI Agents and Hybrid Automation | Saylor Academy | Saylor Academy Unit Readings | CC BY 4.0
OUTPUT LICENSE: CC BY 4.0
5.3: Adding Decision-Making to Agents
This reading supports Unit Learning Outcome 5c.
As you saw in Unit 5.2 when you built a simple AI agent, a useful agent does more than produce a single answer. It must decide what to do next, choose between options, and carry out a sequence of steps in a reliable order.
What decision-making means in an agent
An agent uses decision-making when it compares choices and selects one action instead of another. In practice, this can be as simple as checking whether a message needs a quick reply or a full follow-up, or as structured as choosing the next step in a task chain.
Decision rule: a clear condition that tells the agent what action to take.
Example: If a form submission is missing a required field, the agent asks for the missing information.
Best Use: Use this when the choice is simple and you can describe it clearly.
Limitation: It cannot handle situations that need judgment beyond the rule you wrote.
Priority order: a set of choices ranked from most important to least important.
Example: An agent first checks for urgent customer issues, then routine requests, then informational questions.
Best Use: Use this when the agent must handle more than one possible action.
Limitation: If your priorities are unclear, the agent may choose the wrong path.
How multi-step task execution works
A multi-step task is a task that needs more than one action to finish. The agent must not only choose an action, but also keep track of what it has already done and what comes next.
Step sequence: the order in which the agent completes actions.
Example: First collect the request, then check the data, then draft a response, and finally send it for review.
Best Use: Use this when one step depends on the result of a previous step.
Limitation: If a step fails, later steps may not make sense until you fix the problem.
State tracking: the agent remembers the current stage of the task.
Example: An agent processing a support ticket knows whether it is waiting for a reply, verifying details, or ready to close the ticket.
Best Use: Use this when the same task may pause and continue later.
Limitation: Without good state tracking, the agent may repeat work or skip a step.
Building on the trigger, action, and condition logic from Unit 2.3, you can now see how agents extend those ideas. A rule-based workflow reacts to one condition at a time, but an agent can use the same kind of logic to decide among several next steps and move through a longer process.
A simple real-world example is a travel planning assistant that helps a user book a trip. It may first ask for a destination, then compare flight options, then check hotel preferences, and then summarize the best plan. Each step depends on the information gathered in the previous step, so the agent must decide when to continue, when to ask for more detail, and when to stop.
Job-ready Tip: Designing agent flows
Write your agent as a short decision tree before you build it. Start with the main goal, list the possible next choices, and define what information the agent must have before each choice. This makes it easier to spot missing steps, unclear rules, and places where human review should happen.
As you read this unit, ask yourself: if your agent must make a choice at each step, what information does it need, and how will you make sure it does not skip an important action or repeat work?
Sources / Further reading
(Authored by Saylor Academy, CC BY 4.0)
Estimated reading time: 5 minutes
5.3: Adding Decision-Making to Agents (covers 5a, 5c)
Artificial Intelligence: Foundations of Computational Agents, 3rd Edition — 1.1 What is Artificial Intelligence? — artint.info (Poole & Mackworth, Cambridge University Press) — cc-by-nc-nd
Artificial Intelligence: Foundations of Computational Agents — 1.8 Review (Agent design considerations) — artint.info (Poole & Mackworth) — cc-by-nc-nd
Artificial Intelligence: Foundations of Computational Agents — 8.1 Representing States, Actions, and Goals — artint.info (Poole & Mackworth) — cc-by-nc-nd
Artificial Intelligent Agent Architecture and Clinical Decision-Making in the Healthcare Sector — PubMed Central (PMC) — cc-by-sa