PrerequisiteAI 104 (Agentic AI Workflow & Execution Design) and AI 201 (MCP) recommended
0 / 5 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor
Learning objectives
Distinguish a Fixed Workflow, a Conditional Workflow, an Agentic Workflow and an Adaptive Agent
Explain what makes an agent adaptive: the next step depends on goal, state, observations, available actions and previous results
Describe an agent task using Goal, State, Observation, Action, Environment and Feedback
Name the Constraints, Bounded Autonomy and Stop Conditions that keep an adaptive agent under human control
Choose one real use case, described with simulated data, to carry through the whole course
Problem and need
8 min
When the path cannot be written in advance
In AI 104 you designed workflows: the steps, their order and the checkpoints were decided by a person in advance. That works well as long as you can write the path down. This course starts where that stops. The core question is: when is a Workflow not enough, so that the agent has to decide the next step itself?
Work scenario (simulated)
A research assistant is asked: "Find out why course sign-ups dropped this term." There is no fixed path. The first data source may show that the drop is only in evening classes, which changes what to look at next. A second source may be missing, so another one is needed. The right next step depends on what was just found, and nobody can list every step before starting.
The next step depends on what the last step returned, so it cannot be fully predefined
The same goal may need different paths on different days or with different data
Some steps only make sense after earlier ones have revealed the situation
A fixed script either breaks on the unexpected or ignores it
Warning
More freedom is not automatically better. An agent that chooses its own next step is less predictable, costs more to run and is harder to review. Use adaptive behavior only where the path truly cannot be fixed, and always with limits and a person who remains responsible.
Key takeaway: When the path cannot be fully written in advance, the next step has to depend on the situation. That is the adaptive part.
Core concepts
10 min
Four stages: Fixed to Adaptive
Between a script and an agent there is a progression. Each stage hands a little more of the "what next" decision to the AI.
From Fixed Workflow to Adaptive Agent
Fixed Workflow: every step and its order is defined in advance
Conditional Workflow: defined steps with branches chosen by explicit rules
Agentic Workflow: a mostly fixed flow with some steps judged by the AI within limited options
Adaptive Agent: the next step depends on goal, state, observations, available actions and previous results
Text equivalent: the four stages run in a line from the most predictable to the most flexible. A Fixed Workflow always follows one path. A Conditional Workflow follows one of several paths chosen by rules written in advance. An Agentic Workflow keeps the overall flow fixed but lets the AI make a few bounded judgments. An Adaptive Agent chooses each next step from the situation it is in.
Comparing the four stages (simulated examples)
Stage
Who decides the next step?
Example
Fixed Workflow
The designer, once, in advance
Every Friday: collect feedback, classify, summarize, draft the report
Conditional Workflow
Rules written in advance
If the message is about billing, send it to flow A; otherwise to flow B
Agentic Workflow
Mostly the designer; the AI judges a few steps within limits
The flow is fixed, but "does this comment need follow-up?" is judged by the AI, and unclear cases go to a person
Adaptive Agent
The agent, based on the goal and what it has observed so far
Goal: "find out why sign-ups dropped"; the agent chooses which data to look at next, within allowed sources and a step limit
Check yourself
A mailbox sends each message to one of three predefined handling flows by category. Every flow has fixed steps. Which stage is it?
Key takeaway: The stages differ in one thing: who decides the next step, and how open the choice is.
12 min
Goal, State, Observation, Action: the vocabulary of an agent
An adaptive agent is easiest to understand as a loop. It has a goal, looks at the situation, picks an action, carries it out and looks at what happened. Six words describe that loop precisely, and the rest of the course uses them every time.
The basic agent loop
Goal: what the agent is trying to achieve
Observe State: what is the situation right now?
Choose Action: which available action fits best?
Execute: carry the action out in the environment
Observe Result: the feedback from the environment; then return to Observe State
Text equivalent: the loop starts from a Goal. The agent observes the current State, chooses an Action, executes it, and observes the result. The result changes the State, and the loop repeats until the goal is reached or a Stop Condition applies.
Six terms, illustrated with a simulated sign-up analysis
Term
Meaning
Example
Goal
The result the agent works toward
Explain the sign-up drop in a short note a manager can review
State
What the agent currently knows and has done
Sign-up counts loaded; drop seen only in evening classes; marketing data not yet checked
Observation
New information the agent receives
The marketing export shows no change in ad volume
Action
Something the agent can do next
Open the class schedule; compare with last term; ask a person for missing data
Environment
The world the agent acts in and can only partly see
The data sources, documents and systems it is allowed to use
Feedback
What the environment returns after an action
A table of numbers, an error, an empty result
Workflow
The next step is mostly predefined
Observation only confirms the step is done
Predictable and easy to review
Adaptive Agent
The next step depends on goal, state, observations, available actions and previous results
Observation changes what happens next
Flexible, but needs limits and review
Check yourself
The agent opens a data file and finds that the "evening" column is empty. In the loop, what is that empty column?
Key takeaway: Goal, State, Observation, Action, Environment and Feedback describe every adaptive agent, however it is built.
Boundaries and control
10 min
Constraints, Bounded Autonomy and Stop Conditions
The more an agent decides for itself, the more important its limits become. Three ideas keep an adaptive agent useful and safe: Constraints define what it must respect, Bounded Autonomy defines how much it may decide alone, and Stop Conditions define when it must stop.
Three controls for an adaptive agent
Control
The question it answers
Example (simulated)
Constraints
What must the agent always respect?
Use only the approved data sources; no personal data; finish within 20 minutes
Bounded Autonomy
What may it decide alone, and what needs a person?
It may read and summarize freely; sending anything outside the team needs human approval
Stop Conditions
When must it stop, whether or not the goal is reached?
Goal reached; 10 steps used; two errors in a row; required data missing; needs a decision beyond its limits
Read-only actions can usually be given more freedom than actions that change or send something
The bigger the impact of a mistake, the earlier a person should be involved
A stop condition is a design decision made before the run, not something the agent decides on its own
Info
Human-AI-Human: a person defines the goal and the limits, the agent chooses its steps inside those limits, and a person reviews the result and stays responsible for it. Later modules add Human Escalation, which is the rule for when the agent must hand a decision back to a person.
Check yourself
Which is a well-formed Stop Condition for an adaptive research agent?
Key takeaway: Autonomy always comes with Constraints, Bounded Autonomy and Stop Conditions that a person sets before the run.
Practice and apply
10 min
Practice: Workflow or Adaptive Agent?
Below are 8 simulated scenarios. For each, choose one: Fixed Workflow, Conditional Workflow, Agentic Workflow or Adaptive Agent, and say who decides the next step. 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)
Case
Scenario
Think about
1
Every Monday: export attendance, calculate rates, paste into a template, send for review, always in this order
Is anything decided along the way?
2
Expense claims under a set amount go to automatic checks; claims above it go to a manager, by a rule
Are the branches and their conditions written in advance?
3
A reviewing flow is fixed, but "is this source reliable?" is judged by an AI against a checklist, and doubtful cases go to a person
Mostly fixed, with a bounded AI judgment
4
Goal: "find why lab bookings fell"; the agent picks which records to open, follows leads and stops at 12 steps or when it needs data it lacks
Who decides the next step? What are the limits?
5
A form letter is filled from a template with a name and date
Is there any real decision?
6
A helper that, given "prepare a briefing on options for the new room booking system", searches, reads, changes its approach when sources disagree, and asks a person when it hits a policy question
How open is the path? Where is the person?
7
Messages are sorted into billing, technical or general by an AI, then follow three fixed handling flows
Where does the AI decide, and what happens after?
8
A nightly job backs up files to a set folder and logs the result
Any need to adapt?
Fixed Workflow: every step and its order is defined in advance
Conditional Workflow: defined steps with branches chosen by rules written in advance
Agentic Workflow: mostly fixed flow, some steps judged by the AI within limited options
Adaptive Agent: the next step depends on goal, state, observations, available actions and previous results
1Apply: choose one real use case for the whole course Pick a task where the path really cannot be fully predefined: the next step depends on what you find. Every later module practices on it, and it ends up in your Agent Planning & Reasoning Canvas.
2Describe it with simulated data Write only the goal and the nature of the work, with no real names, client personal data or company secrets. If the task involves sensitive content, use a simulated version.
3Write the Goal and the limits One sentence for the Goal, for example "Explain why sign-ups dropped, in a note a manager can review", plus one Constraint and one Stop Condition.
You need to pick one use case for the whole course. Which is the best fit?
Key takeaway: Human-AI-Human: a person defines the goal and limits, the agent chooses its steps inside them, and a person reviews and stays responsible for the result.
