Assistance, Workflows and Agents
The same word, AI, covers a single reply, a fixed sequence of steps and a system that chooses its own next move. For teaching, the differences that matter are how much of the path the system chose and how much of it you can see afterward.
An instructor wants six short items that test the difference between negative reinforcement and punishment, two ideas students can confuse. She asks a tool for items and an answer key, reads them and edits. That is one exchange. In a second arrangement she gives a system the goal of building a week of practice on operant conditioning, and the system decides which readings to consult, which tools to run, in what order and when to stop. The label AI covers both. For teaching, what separates them is not how intelligent either is. It is how much of the path was chosen by the system, and how much of that path she can inspect afterward.
Three arrangements
Ordinary drafting assistance is a request and a reply, possibly repeated. A predefined multi-step workflow is a fixed sequence that someone designed, such as retrieve passages, draft items, format them. A model may do the work inside each step, but the sequence itself does not change. An agent, as the word is used in these guides, is a system in which the model itself chooses which actions to take and which tools to use next, based on what it finds. The boundaries are graded, and one product can behave one way on one task and another way on another. A fixed pipeline is not agentic simply because it has many steps. A single reply is not harmless simply because it has one.
What each arrangement can hide
Drafting assistance affects more than wording. It chooses which scenario illustrates punishment, which verbs carry the idea and sometimes which source to cite, and the instructor's edit begins from those choices. Pipelines and agents can add steps the instructor does not directly observe. Whether those steps stay invisible depends on design. A system that records its plan, the sources it opened and the checks it ran leaves a trail a person can review. One that returns only a final text does not. Being agentic does not by itself hide decisions, and being simple does not guarantee they are visible. The useful question to ask of any arrangement is what trail it leaves.
Return to the items. For each one, the instructor needs to know what scenario it uses, whether the key is right, and whether it tests the intended distinction or a cue in its wording. The classic slip is the one students make too. Something unpleasant is removed and the behavior increases, which is negative reinforcement. Something unpleasant is added and the behavior decreases, which is punishment. The label follows what happened to the behavior, so a scenario that never says whether the behavior later rose or fell cannot be keyed reliably. Reading a trail, where one exists, shows where a slip entered. Without a trail, the instructor has only the finished text.
The same questions apply to someone studying alone. Asking a tool to explain the two concepts and then testing that explanation against a textbook keeps the checking close at hand. Delegating a multi-step plan, such as find readings, write notes and generate practice questions, may shift many choices at once. Delegation can be reasonable, and which choices a learner wants to keep depends on the goal. If the goal is to tell the two concepts apart in a scenario never seen before, the choices about what to practice and how to check are probably worth keeping. That is an expectation to examine in a course, and no established result is claimed for it.
An original classroom example
Original, unevaluated example. Version A: the instructor asks a tool for six negative reinforcement and punishment items with keys, then edits them. Version B: a fixed pipeline retrieves textbook passages, drafts items and formats them. Version C: a system chooses its own sources and steps toward the goal of a week of operant conditioning practice. For each version list the decisions about scenarios, sources, wording and keys, say who made each, and say what trail the instructor would have to check a wrongly keyed item.
Questions for your course
- For one task you prepare, which choices does a drafting tool already shape before you edit?
- What trail would let you review what a multi-step system did?
- How would you tell a fixed pipeline from a system that picks its own steps?
- Which choices would you keep if the goal were learning to tell two concepts apart?