Assistant or agent?

In the last lesson you read the transcript of a toy agent and named each turn of its loop. This lesson steps back and asks a question people get wrong every day: is the product in front of you an assistant or an agent? Vendors sell both with the same words, and both often show a chat box. By the end you have one test that answers the question whatever the interface looks like, and you know which part of an agent to look at when it does something you didn’t expect.

The model and the software around it

A model takes text in and gives text out, once. It has no hands. It can’t open your inbox, read a file, or send anything. When you type into a plain chat assistant with no tools and get a reply, the product made one call to the model and showed you the reply, and now it waits for you [1].

An agent is software that calls the same kind of model again and again. Before the first call it tells the model which tools it has. When a reply asks for a tool, the software runs it, puts the result back into the conversation, and calls the model again. It keeps going until the model writes an answer instead of a request, or until a stop rule ends the run [1]. You saw that loop turn by turn in the last lesson, and model vs agent is the name for the difference it makes: the model supplies the judgment, and the software supplies the tools, the memory and the loop.

An assistant and an agent look alike from the outside because both start from a prompt and end with text. The difference is what happens between the prompt and the text. An assistant produces text for a person to act on. An agent acts, through its tools, and decides on its own what to do next [2].

One test: what can it do without another prompt?

The interface tells you nothing. A chat window can front an agent, and a big button labeled “Automate” can trigger one plain model call. Many chat products now have tools such as web search, so the window you type into says nothing about which kind you have. So ask the question that the interface can’t hide: after you give it a task, what can this system do before it needs you again?

• If the answer is “write text, and then wait for me”, it is an assistant. Everything that happens to the world afterwards happens because a person did it.
• If the answer includes “look something up”, “change a file”, “send a message” or “run the next step because the first one worked”, it is an agent. Some of what happens to the world happens because the software did it.

Picture two products that both show a chat box. You paste a question about last quarter’s sales into each. The first gives you a database query to run. The second runs the query, reads the numbers that come back, notices one region is missing and runs a second query for it. Same box, same kind of model, different answers to the test. The first is an assistant and the second is an agent, and the second one had a tool that reached your database [3].

The test is also a habit for reading a product page. Ignore the words “assistant”, “copilot” and “agent” in the name, because vendors use all three for both kinds. Look for the verbs that describe what the product does after your prompt. “Drafts”, “suggests” and “summarizes” describe text. “Books”, “sends”, “updates”, “runs” and “keeps going until” describe actions.

Checkpoint · sort

Assistant or agent?

Unplaced (drag an item to a bucket, or select it and then click the bucket)

A chat window that drafts an email reply for you to copy and sendA tool that reads your inbox, drafts replies and sends them without another promptA button in a spreadsheet that explains the selected formula in plain wordsA chat window that runs your database query, reads the result and runs a follow-up query on its ownA support product that searches your help center, tries a fix on the customer account, checks whether it worked, and closes the ticket

Assistant

Agent

CheckHintSkip

The harness

The software around the model has a name: the harness. It is the part that makes a model into an agent, and it has more in it than the loop you watched. The Agent Engineer Course calls it the orchestration layer, and lists the loop, the state, error handling, stop conditions and guardrails as its jobs [1]. A harness holds:

• The loop. Call the model, read the reply, run the tool it asked for, feed the result back, repeat.
• The tool definitions. The text that tells the model which tools exist, what each does and what parameters it takes. In the last lesson this was the two-line description of list_dir and read_file.
• The permission checks. The step between a tool request and the tool, where the harness can ask a person, refuse, or run.
• The system prompt. The standing instructions the model reads before every task.
• Context management. What the harness keeps in the conversation, what it drops when the conversation grows, and what it saves for the next run.
• The stop rules. The step limit, the “model said done” check, and the stop button.

Two agents built on the same model can behave very differently, because their harnesses differ. One has a file-write tool and the other doesn’t. The first asks before every change, where the second asks only before a deletion. To save space, the first drops old tool results from the conversation, and the second keeps everything. None of this is the model’s doing, and a person who compares “which model does it use” is comparing the one part the two products share.

The same fact points at where to look when an agent misbehaves. An agent that sent an email it shouldn’t have had an email tool and a permission check that let it through. An agent that forgot the instruction you gave it twenty steps ago hit its context management. An agent that ran for an hour and produced nothing had no limit low enough to stop it. In each case the fix is in the harness: take a tool away, or add a check or a stop rule. The model’s judgment matters too, and a better model helps with a wrong decision, but the reach and the guardrails are the harness’s.

Checkpoint · multi-choice

Which of these are decided in the harness?

Two agents run on the same model. Which of these differences between them are set in the harness?

Select exactly 3.

Whether the agent can send an email at all.Whether the agent asks a person before it deletes a file.Whether the agent reads a clear paragraph and understands it.The number of steps the agent gets before the run ends.Whether the agent gives a confident answer that turns out to be wrong.

CheckHintSkip

List the tools before you judge the reach

Once you know something is an agent, the next question is what it can reach, and the answer is its tool list. A research agent with a web search tool can read the web. Give it an email tool as well and it can write to other people. The two are sold under the same name and have very different consequences when something goes wrong.

So read the product page, or the settings screen, and write down the tools before you form an opinion. A list like “web search, file read, calendar write” tells you which one acts on others: calendar write is the one that puts something on a colleague’s screen. A list like “read the inbox, draft, send” tells you that a wrong draft can reach a customer. Judging an agent without its tool list is judging a car without knowing whether it has brakes.

Checkpoint · choice

Which tool changes what others see?

A product page for a meeting assistant lists its tools as “web search, file read, calendar write”. Which tool deserves the closest look before you turn the product on, and why?

Web search, because it reaches outside the company.File read, because your files hold confidential material.Calendar write, because it puts an entry on a colleague's calendar.

CheckHintSkip

Pitfall: A chat window with an email tool

A manager tries a new “AI assistant” that comes with the company’s mail program. It opens as a chat window that looks like the drafting assistant she already uses, and she treats it the same way. She types “reply to the supplier and tell them the delivery date is fine”, expecting a draft to read and correct. The reply appears in the chat, and under it a line she doesn’t read: “Sent to supplier”. The product had a send tool, and its harness had no check before a send. The window looked like the old assistant, and the test would have caught it in a minute: after one prompt, what can this do without me? The first time you meet a chat box, find its tool list before you type anything you don’t want acted on.

Exercise

Find five products or features that you or your team use or have been offered, each described with AI in some way. A mail program’s assistant, a coding tool, a meeting notetaker, a search feature, and a customer support bot are common finds. For each one, write two lines in a note. The first line answers the test: after one prompt, what can it do without another prompt from you? The second line either says “assistant” or says “agent” followed by the tools you can find on the product page or the settings screen. Twenty minutes is enough.

In a good result, the classification of each product comes from what the product does, never from its name. Every “agent” line has a tool list, even a short one. If you couldn’t find the tool list for a product you marked as an agent, write that down too, because it is the most useful thing you learned about that product. Which of the five surprised you, and was the surprise about the interface or about the tools?

Stretch: For each product you marked as an agent, name the one tool you would remove, or put a permission check in front of, before letting a colleague use it on real data.

Recap

1. A model takes text in and gives text out, once. An agent is software that calls the model in a loop, gives it tools, feeds the results back and stops when the goal is met [1].
2. Tell the two apart by what the system can do without another prompt, whatever its interface looks like. A chat window can front an agent, and a button can trigger one plain model call.
3. The harness is the software around the model: the loop, the tool definitions, the permission checks, the system prompt, context management and the stop rules. Agents on the same model differ because their harnesses differ, and the fix for a misbehaving agent is most often in the harness.
4. List an agent’s tools before you judge what it can do. What it can touch is what its tools can touch, and the tool that acts on other people is the one to look at first.

You can now

• Tells a chat assistant from an agent by what it can do unprompted

Mark lesson finishedI know this, skip it

References

1. Addy Osmani, Ivar Soares Urdalen, Leo Simons. What are AI agents: model versus agent, autonomy levels, when a prompt suffices. Agent Engineer Course. Course. AEC-01
2. DeepLearning.AI. Agentic AI: M1 workflows and autonomy, M2 reflection, M4 evals and error analysis, M5 autonomous agents. DeepLearning.AI. Course. DLAI-11
3. Anthropic. Claude Platform 101. Claude Academy. Course. Academy claude-platform-101