AI agent vs AI assistant: what the difference means for a small team
An AI assistant answers one message at a time with your access. An AI agent takes multi-step actions under a brief and granted permissions.
On this page
An AI assistant answers. You send a message, it replies, and nothing changes unless you act on the reply. An AI agent takes action. You give it a goal, it plans steps, calls tools to read and change things, checks results and continues until the task is done or it needs you. The difference is not intelligence; it is that an agent can change your records, so it needs a brief, permissions and a record of what it did.
That distinction decides what to buy, who may switch it on, and what can go wrong. This post defines both, shows the difference with concrete examples from a small team's week, gives a checklist for deciding which one a task needs, and is specific about how Notion, Linear and Atrix have drawn the line as of October 2026.
What an AI assistant does
An assistant maps one message to one reply. The model reads what you give it, which may be a question, a pasted document or an open page, and produces text: an answer, a draft, a summary, a translation, a list of options. Then it stops.
Three things follow from that shape:
- It acts as you. Whatever it can see is whatever you can see. There is no separate permission to think about, because it cannot change anything on its own.
- You are the loop. If the task needs four steps, you ask four times and carry the results between them.
- Nothing is left behind. The output is in the chat. If it matters, you copy it somewhere.
This is not a weakness. For writing, thinking, explaining and first drafts, an assistant is the right tool, and most people on a team use one daily. The limits only show when the work is operational: file these, update those, find out what is overdue and tell the right person.
What an AI agent does
An agent runs a loop. Given a goal, the model decides a first step, calls a tool to search, read or write something, looks at the result, and decides the next step. It keeps going until the goal is met, it hits a limit, or it reaches a point where it must ask a person. The glossary entry on AI agents goes into the failure modes; the short version is that an agent is software that pursues a goal over several steps and can change things along the way.
Because it can change things, an agent has parts an assistant does not:
- A brief. A goal with enough context to plan against. "Prepare this project for launch" is a brief; "hi" is not.
- Tools. Functions it may call, each with a defined input and output. The Model Context Protocol has become the usual way to give an agent tools across products.
- Permissions. What it may read and what it may write, ideally narrower than what the person who started it can do.
- Checkpoints. Points where it must stop and ask: before sending, before deleting, before spending.
- A run. A bounded unit of work with a start, an end, a status and an output. The agent run entry explains why vendors meter by runs rather than messages.
AI agent vs AI assistant: side by side
| AI assistant | AI agent | |
|---|---|---|
| Unit of work | One message, one reply | One run: goal to result over several steps |
| Can change records | No | Yes, within granted permissions |
| Who carries context between steps | You | The agent |
| Acts as | You, with your full access | A participant with scoped permissions |
| Stops to ask | Never needs to | At defined checkpoints |
| Leaves behind | Text in a chat | Documents, issues, records, and a log of the run |
| Typical failure | A wrong or made-up answer | A wrong action, repeated, on stale context |
| Governance needed | Little | Permissions, audit, retry and spend limits |
| Good for | Writing, thinking, questions | Planning, filing, tracking, follow-up |
Read the "typical failure" row twice. An assistant's mistake costs you a few minutes. An agent's mistake can file twelve duplicate issues or email a customer, which is why every other row in the agent column exists.
Four examples from a small team's week
We are a company of a handful of people running on these tools, so these are the shapes of task we actually hand to each.
Monday: a product brief. A founder has two paragraphs describing a customer portal. An assistant will turn them into a decent requirements list in one reply, and the founder pastes it into a document. An agent given the same brief reads the project's existing pages, writes the requirements document into the project, proposes an architecture note, and adds four planned issues to the backlog for review. The founder edits two requirements and deletes one issue. Same model quality; different amount left behind.
Tuesday: "what is overdue?" An assistant can only answer if you paste the issue list. An agent with read access to the tracker lists the overdue items by owner and, if it has permission, posts a short update in the channel. Without permission to post, it drafts the update and stops.
Wednesday: a customer asks for a feature. Someone pastes the email to an assistant and asks how to reply; it drafts a reply. An agent with CRM and work access finds the customer's company, attaches a note to the deal, checks whether an issue already covers the request, and files one if not, linked to the deal. It does not send the reply; sending is a checkpoint.
Friday: cleaning up the backlog. An assistant can suggest which of twenty issues look like duplicates if you paste them. An agent can read the backlog itself and propose merges, but this is the task where a checkpoint matters most: merging is hard to undo, so a well-designed agent proposes and a person confirms.
The pattern: an assistant is faster for a single answer; an agent is worth it when the answer needs to become records, and only when someone has decided what it may touch.
How three products draw the line
Vendors do not agree on where assistant ends and agent begins, so it is worth reading how three of them have drawn it.
Notion shipped Custom Agents in Notion 3.3 on 24 February 2026, after announcing them around Notion 3.0 in September 2025. Its release notes describe agents that run on a schedule or a trigger without a prompt each time, with permissions set like a teammate's, every run logged with what triggered it and what it changed, and changes reversible. That is the agent shape: brief, permissions, run, log. The earlier Notion AI, which answers in a sidebar, is the assistant shape, and both exist in the same product. Custom Agents were free until 3 May 2026 and use credits after that.
Linear took a different route: it does not ship its own agent but lets third-party agents join a workspace. Its agents documentation says issues are assigned to humans and delegated to agents, so a person always remains accountable for an issue an agent is working on. Its Agent Interaction Guidelines ask agents to disclose that they are agents, give instant feedback and be transparent about their state. The agent APIs are in developer preview. The human-stays-assignee rule is the clearest statement we have seen of the governance an agent needs that an assistant does not.
Atrix is our own product, so weigh this accordingly. The agent works in run modes on a project: Develop writes requirements and architecture from a brief, Market writes positioning and a launch plan, and the requirements audit freezes a baseline and computes coverage from records rather than from the model's opinion. It proposes two to five issues at a time into the backlog, never straight into a plan, and never pushes code. Runs are metered, 100 a month on Builder and 500 on Studio, and chatting with the agent uses none, which is the assistant and agent split in one price. Outside agents can act in the workspace over MCP with OAuth 2.1 scopes that never exceed the person's own permissions. Limits: the Scale run mode is not launched, and the agent does not run on a schedule the way Notion's do. The AI agent page lists what each run produces.
Which one does this task need?
A short checklist. Answer yes to any of the first three and you want an agent; answer yes to the last two and you want an assistant.
- Does the result need to exist as a record such as an issue, a document in the project, a note on a deal, rather than text you will copy?
- Does it take more than one step with the output of one feeding the next?
- Would you hand it to a junior colleague with a written brief and expect them to come back with it done?
- Is the task thinking, writing or explaining that ends when you have read the answer?
- Would a wrong answer cost minutes, not hours, because nothing changed?
If you want an agent, decide three things before switching it on: what it may read, what it may write, and where it must stop and ask. If nobody on the team will make those decisions, use an assistant and keep the loop yourself.
FAQ
What is the difference between an AI agent and an AI assistant?
An assistant answers one message at a time using what you show it and cannot change anything on its own. An agent pursues a goal over several steps, calling tools to read and change records under permissions someone granted, and leaves a log of what it did.
Is a chatbot an AI agent?
No. A chatbot is an assistant: it replies to messages. It becomes an agent only when it can call tools, take multi-step actions and change records, with permissions and a record of each run. "Agentic AI" is the usual name for that step.
What is an AI agent at work?
Software that, given a brief, does operational work across a team's tools: reads the project, writes documents, files issues, updates records, and stops at checkpoints for a person to approve. At work the permissions and audit matter more than the model.
Are AI agents safe to give access to company data?
As safe as the permissions you grant and the checkpoints the product enforces. Grant read access first, write access only where the writes are reversible, and keep a person on anything hard to undo. Linear's rule that a human stays accountable for every delegated issue is a good model.
What does an AI agent cost compared with an assistant?
Assistants are usually priced per seat or included in a plan. Agents are often metered by runs or credits because each run can call many tools. Notion's Custom Agents use credits after 3 May 2026; Atrix includes 100 or 500 runs a month depending on plan and does not charge runs for chat.
Next step
The AI agent and agent run glossary entries have the definitions and the questions to ask a vendor. To see what each run mode leaves behind in Atrix, and what it does not do, read the AI agent page.