Agentic or Not

Eight AI setups. For each one, decide whether it is agentic, partly agentic or not agentic at all — then compare your reasoning with ours.

“Agentic” gets used for almost anything an AI does lately, which makes it hard to tell when the word is carrying real weight. The question worth practicing is narrower than the label: which parts of the setup are actually doing the work, and who decides what happens next.

For this activity, use this working definition. An agent is a model working toward a goal in a loop: it takes steps, uses tools, checks what happened, and decides what to do next.

That is a useful frame rather than a settled industry definition. Agentic behavior sits on a spectrum, and reasonable people draw the line in different places. You will probably disagree with at least one of our calls, which is the interesting part.

8 scenarios one decision each 5–8 minutes No account, nothing saved

The eight scenarios

Each one gives our call and the reasoning behind it. If you are reading this list rather than working through the scenarios one at a time, the activity needs JavaScript — the lesson is all here either way.

Clear cases

You paste a question into a chatbot (“What was the holding in Palsgraf?”) and read the answer.

Our call: Not agentic. The model answers once and stops. Nothing else in the loop is involved, since there are no tools and no further steps, and every decision about what happens next is yours.

A custom chatbot that has standing instructions (“Always answer like a patient torts professor”) but no tools or files.

Our call: Not agentic. Standing instructions shape how the model answers, but it still answers once per message and takes no steps of its own, so instructions by themselves don’t make it an agent.

A Python script that reads every PDF in a folder of Supreme Court opinions and lists the dissents one justice wrote or joined.

Our call: Not agentic. There’s no model here at all. The script is a deterministic program that does exactly the same thing every time and never decides anything, and that predictability is the reason to use it.

The hard middle

You upload a case and say “brief this case.” The AI notices it has a case-brief Skill installed, follows its steps, and returns a brief in the Skill’s format.

Our call: Partly. The model chose a Skill and used your file as context, so more than the model is at work. It still makes a single pass, though, without taking further actions or checking its result before it stops.

An assistant drafts a reply to a client email and leaves it in your drafts folder for you to review and send.

Our call: Partly. It uses a tool, your email, in pursuit of a goal, but the design keeps a human checkpoint before anything is sent. For legal work, that is often exactly the right arrangement.

The SLS Writing Partner reviews your memo: it runs the AI Use Gate, picks the claims and source traceability check, works through each citation, and flags what needs human verification.

Our call: Partly. Its instructions, Skills, and your file work together over several steps. By design, though, it hands every decision back to you, which keeps it short of the full loop.

See how the Writing Partner is built.

The fuller loop

Asked to “find this week’s new state AI laws and update our tracker,” an assistant searches the web, opens sources, compares them with the spreadsheet, edits rows, and rereads its changes before it reports back.

Our call: Agentic. This one has a goal, several tools, and a loop in which it acts, checks the result, and decides what to do next. That is the full agent loop, which is also why its work needs the closest supervision.

Asked to “fix the failing test,” a coding assistant runs the tests, reads the error, edits a file, and reruns the tests until they pass.

Our call: Agentic. This is the loop in its textbook form. The assistant tries something, observes what happened, adjusts, and repeats until the goal is met, choosing each next step itself.

Where did you draw the line differently, and why?

The scenarios in the hard middle are the ones worth arguing about. Each of them involves more than a single model answering a single question, and each of them still hands the next decision back to a person. Whether that is enough to call something an agent is a judgment, not a fact.

If you want to go further into how agents are actually assembled, Agents Without the Hype takes the loop apart over ten modules.