# What Are Notion AI Agents? Everything They Can Actually Do

> Canonical HTML version: https://thejayant.in/blog/notion-ai-agents-explained
> Author: Jayant Solanki — https://thejayant.in/
> This Markdown file is a plain-text twin of the article at the URL above. Same content, no page furniture. It is public, not bot-only.

- **A Notion AI agent is software that does multi-step work inside your workspace** — creating pages, updating hundreds of database rows, reading Slack and the web — rather than just writing text for you to paste.
- **There are two kinds.** The personal _Notion Agent_ runs when you ask it, for over 20 minutes at a stretch. _Custom Agents_ run on triggers and schedules without you.
- **They are not automations.** Automations are free, deterministic if/then rules. Agents cost credits and use judgement. Choosing wrong is the main source of wasted Notion AI spend.
- **Access is opt-in, not default.** An agent can only touch pages, databases and apps you explicitly grant. Notion is clear that agents "never have full workspace access by default".
- **Cost:** Business or Enterprise plan, plus Notion credits — free to try, then $10 per 1,000 monthly credits.
- **The best-documented result so far:** James Lawley, IT Ops Manager at Remote, saved his team **20 hours a week** with a request-routing agent.

Agent announcements are hard to read from the outside. Every vendor says theirs is autonomous. Most of them mean "it writes a draft". This is the agent chapter of [the full guide to Notion for productivity, AI and automation](https://thejayant.in/blog/notion-productivity-ai-automation-guide), taken properly apart.

This is what Notion's agents actually do, what they cost, where they fail, and how to set one up — with every capability traced to Notion's own documentation rather than its marketing. The workspace layer underneath — databases, relations and automations — is covered [in the pillar guide](https://thejayant.in/blog/notion-productivity-ai-automation-guide).

## What is a Notion AI agent?

**A Notion AI agent is an AI system that performs multi-step work directly inside a Notion workspace — reading pages and databases, creating and editing content, and pulling context from connected tools like Slack, Google Drive and the web — acting only within permissions you explicitly grant it.** Unlike a chatbot that returns text, an agent changes the workspace itself.

Notion rebuilt its AI as agents in [Notion 3.0](https://www.notion.com/releases/2025-09-18), released 18 September 2025. The framing in that announcement is the useful one: with 1.0 and 2.0 you got tools to do your work in one place; with 3.0 you get agents that do the work for you.

## Personal Agent vs Custom Agents vs automations

Three things get called "Notion AI" and they behave completely differently. This table is the one to internalise.

|  | Automation | Notion Agent (personal) | Custom Agent |
| --- | --- | --- | --- |
| **What it is** | A deterministic if/then rule on a database | Your personal assistant, run on demand | A configured agent that runs itself |
| **Starts when** | A trigger fires | You ask it | A trigger or schedule fires |
| **Uses judgement** | No | Yes | Yes |
| **Costs credits** | No | Yes | Yes |
| **Shared with the team** | Lives on the database | No, it is yours | Yes, with permission levels |
| **Runs while you sleep** | Yes | No | Yes |
| **Plan** | Free (Slack action on paid) | Business | Business or Enterprise |
| **Best for** | Fixed logic: set a date, send a notification | One-off multi-step jobs you supervise | Recurring work: triage, standups, reports |

**If you can describe the rule in an if/then sentence, use an automation.** It is free, instant and identical every time.

**If it needs reading and interpreting, use an agent.**

**If it needs an external API or exact arithmetic, use a Worker** and let the agent call it as a tool.

Most Notion AI overspend traces to a single mistake: paying an LLM to do something a formula or an automation does exactly, free, forever.

## What agents can actually do

Five capabilities, each from Notion's own documentation rather than a demo video.

### 1. Run for over 20 minutes on one task

Notion states the personal Agent is "capable of over 20 minutes of multi-step actions with a state-of-the-art memory system", using Notion pages and databases as that memory. This is the capability that separates an agent from a chat completion — it can hold a plan across many steps and keep going.

### 2. Work at database scale

Per the 3.0 release notes, agents work "at a scale no person could — updating or creating hundreds of pages at once". A restructure that would take an afternoon of manual editing becomes one instruction.

### 3. Pull context from connected tools

Agents read from Notion pages and databases and from connected apps — Slack, Google Drive, GitHub and others — plus the open web if you turn web access on. Web access is a toggle per agent: on when the workflow genuinely needs outside information, off for strictly internal work.

### 4. React to events, not just schedules

Custom Agents trigger on recurring schedules (daily, weekly, monthly, weekends only, at a specific time and timezone) and on real events: a property updated in a database, a page created or removed, **an AI Meeting Note finishing**, an emoji reaction added to a Slack message, or the agent being @-mentioned. Triggers can be combined and filtered — only this database view, only when status is X, only when a message contains a keyword.

### 5. Hand work off to other agents

A Custom Agent can pass part of a job to other Custom Agents mid-run, each with its own instructions, context, access and model. Notion's own example: a reporting agent hands off planning, data prep and review, then assembles the final report itself.

This is also a cost lever. Notion's guidance is to give the primary agent a stronger model for planning and assembly, and give focused sub-agents lighter, faster ones — keeping most of the work cheap and saving deep reasoning for the steps that need it.

## Five jobs agents are genuinely good at

Ranked by how reliably they pay back, based on running them across client accounts.

1. **Triage.** Reading an unstructured inbound stream — support tickets, form submissions, emails, Slack requests — and turning it into structured rows with an owner, a category and an urgency flag. This is the highest-ROI category and the one behind the Remote result below.
2. **Meeting follow-ups.** An AI Meeting Note finishes, which triggers an agent that extracts commitments, infers owners and due dates, and writes them as linked task rows. A 45-minute call becomes a filled task list.
3. **Standups and status reports.** Read the projects database on a schedule, summarise what moved, flag what has not, post to Slack. Nobody has to chase anybody.
4. **Report drafting.** Read the numbers already on a page and draft the narrative. Reviewing a draft takes minutes; writing one from scratch takes half an hour.
5. **Database cleanup.** Find rows with empty required properties, duplicate titles, or past-due dates with a live status. Ask for a list, not an edit.

Notion's published reference point for the ceiling: in the [3.3 release notes](https://www.notion.com/releases/2026-02-24), **James Lawley, IT Ops Manager at Remote, saved his team 20 hours per week** with a request-routing agent. Notion's own security team reports a further 6+ hours a week from a similar one. Both are triage.

## Three jobs agents are bad at — and what to use instead

**1. Anything deterministic.** Setting a date when a status changes, calculating a score, posting a fixed message. Use an automation or a formula. You are paying credits for a worse version of something free.

**2. Exact arithmetic and API calls.** Pulling Search Console numbers, summing revenue, hitting an external endpoint. An LLM reasoning its way to a number is slower, more expensive and occasionally wrong. Use a Notion Worker and register it as a tool the agent can call.

**3. Final client-facing copy.** Agents over-assign, over-claim and write confidently about things they half-know. Use them to draft into a staging database and promote after review — never straight into a shared wiki or a client deliverable.

There is a fourth, quieter one: **work where nobody has written down the process.** An agent pointed at a well-written SOP database gives dramatically better output than one guessing at how your team does things. If the process only exists in someone's head, write it down first — then the agent is worth building.

## What agents cost, and how to keep the bill sane

Two separate costs, which is the part people miss.

- **The plan.** The personal Notion Agent needs **Business** ($20 per member per month). Custom Agents need **Business or Enterprise**.
- **The usage.** Custom Agents and Notion Workers both consume **Notion credits** — free to try, then **$10 per 1,000 monthly credits**, billed on usage rather than per agent. Both appear in the same credits dashboard.

Four things that actually move the bill:

1. **Scope.** An agent given three databases costs less and performs better than one given the whole workspace. Broad scope is the single biggest source of waste.
2. **Model choice.** Custom Agents run on different models — Claude Fable 5 (Business and Enterprise only), Claude Sonnet 5, the newest GPT models, Gemini and Grok. **Auto is the recommended default.** Premium models spend credits at their own price.
3. **Hand-offs.** Sub-agent work consumes credits too, so extra hand-offs cost more — especially on a premium model. Start with one or two sub-agents.
4. **Deterministic work.** See above. Move it to automations and Workers.

Check the **Activity** and **Insights** tabs after a few runs and compare model use, output quality and credits across them. That comparison is where the tuning actually happens. The same logic applies to any AI workload — I have worked it through generally in [cost per task, not cost per token](https://thejayant.in/blog/gpt-6-astra-cost-per-task).

## Safety: scope, approval and model control

This is the section to read before connecting anything to a client workspace.

### Agents only see what you grant

Notion is explicit: agents "act only on the pages, databases, and external apps you explicitly grant access to. They never have full workspace access by default." You configure this per agent under **Tools and access**.

One trap worth knowing: **linking to a Notion page inside your instructions does not grant access to it.** The page has to be added in Tools and access. An agent that seems to be ignoring a document it was told to read is usually hitting exactly this.

### Permission levels

| Level | Can do |
| --- | --- |
| **Full Access** | Configure instructions, triggers, access and models; view and manage activity logs; run and interact |
| **Can Edit** | Modify instructions and configuration; review activity |
| **Can View and Interact** | Run and chat with the agent; view Settings read-only. Cannot edit or share |

Note the edge case in Notion's docs: people _without_ access can still trigger an agent configured to respond to events such as Slack messages in channels they can reach. If an agent is wired to a public channel, treat everyone in that channel as able to invoke it.

### Review, reversibility and audit

- **Every run is logged.** The Activity tab records what triggered a run, the actions taken, and any errors — visible to Full Access users.
- **Agent configuration has version history**, so a bad instruction change can be rolled back.
- **Enterprise workspaces get audit-log visibility** on instruction updates, permission changes and integration additions.
- **Turn on line-by-line edit approval** for any agent touching client-facing content or a shared wiki.

### The model-control decision agencies should make on day one

Notion documents this plainly: when you use Claude Fable 5, **Anthropic may store your prompts and responses for a period of time** to keep the model safe and working well. The other models available in Notion do not work this way. That is precisely why an admin has to turn it on before anyone in the workspace can select it.

If you handle client data, this is a decision to make deliberately at **Settings → Notion AI → General → Allowed models for Custom Agents**, not one to discover during a security review. Models an admin turns off appear dimmed, and any agent using a disabled model automatically switches to your default or to Auto.

## Set up your first agent in ten minutes

Use desktop or web — Custom Agents cannot be built, edited or used on mobile.

1. **Pick a triage job.** Something with messy input and a structured output. Not "help me write".
2. **Build the destination first.** Create the database the agent writes into, with the exact properties you want filled. An agent with a defined output shape produces usable records; one told to "write a summary somewhere" produces sludge.
3. **Open the Agents section in the sidebar** and select **+**. You can describe what you want in natural language and let Notion draft the instructions, triggers and access for you to review — or start from an empty instructions field.
4. **Write the job description properly.** Start with the job and the outcome, then add concrete steps, inputs and outputs. Compare: _"Summarise yesterday's support tickets"_ against _"Group yesterday's tickets by product area, list the top 3 recurring issues with ticket counts, flag anything mentioning billing or data loss as urgent, and write it as five bullets max in the Support Digest database."_
5. **Set the trigger.** A schedule with an explicit timezone, or a Notion or Slack event with filters.
6. **Set access narrowly.** Two or three databases. Web access off unless the job needs it.
7. **Leave the model on Auto** for the first week.
8. **Test in the Chat tab** before you publish. Ask it to run the job once and read what comes back.
9. **Watch the first five runs** in the Activity tab and tighten the instructions after each one.

Treat configuration as a conversation, not a form. Use the agent's own Chat tab to test instructions, feed it the errors from Activity, and iterate — the same way you would brief a new hire, notice what they misunderstood, and clarify.

Agents that get written once and published never get good. Agents that get five rounds of tightening in week one usually do.

## FAQ

### What is a Notion AI agent?

A Notion AI agent is an AI system that performs multi-step work directly inside a Notion workspace — reading pages and databases, creating and editing content, and pulling context from connected tools like Slack, Google Drive and the web — acting only within permissions you explicitly grant it. Unlike a chatbot that returns text for you to paste, an agent changes the workspace itself.

### What is the difference between Notion Agent and Custom Agents?

The personal Notion Agent runs when you ask it, works on one job at a time under your supervision, and is not shared. Custom Agents are configured once with a job, a trigger or schedule, a data scope and an output destination, then run automatically in the background without you starting them. Custom Agents can also be shared with your team with permission levels, and can hand parts of a job off to other Custom Agents.

### Do Notion agents cost extra?

Yes, twice over. The personal Agent requires the Business plan at $20 per member per month, and Custom Agents require Business or Enterprise. On top of the seat cost, Custom Agents consume Notion credits — free to try, then $10 per 1,000 monthly credits, billed on usage rather than per agent. Notion Workers draw on the same credit pool.

### Can a Notion agent access my whole workspace?

No, not by default. Notion states that agents act only on the pages, databases and external apps you explicitly grant access to, and never have full workspace access by default. You configure this per agent in its Tools and access settings. Note that mentioning a page inside the instructions does not grant access to it — the page must be added explicitly.

### How long can a Notion agent run?

Notion states the personal Agent is capable of over 20 minutes of continuous multi-step action on a single task, using Notion pages and databases as its memory system. Custom Agents run on triggers and schedules rather than a single session, so they continue operating in the background indefinitely once published.

### What triggers can a Notion Custom Agent use?

Recurring schedules — daily, weekly, monthly, yearly, weekends only, at a specific time and timezone. Notion events — a page created in a database, a property updated, a page removed, or an AI Meeting Note finishing. Slack events — an emoji reaction added to a message, or the agent being @-mentioned. Multiple trigger types can be combined on one agent, and filters can restrict runs to a specific database view, property value or keyword.

### Which AI models can Notion Custom Agents use?

Supported models include Claude Fable 5 (Business and Enterprise plans only), Claude Sonnet 5, the newest GPT models, Gemini and Grok, with Auto as the recommended default. Workspace admins can restrict which models are allowed; disabled models appear dimmed and any agent using one automatically switches to the default or to Auto. Claude Fable 5 requires an admin to enable it first because Anthropic may store prompts and responses for a period, unlike the other available models.

### Are Notion agents safe to use with client data?

They can be, with three settings configured deliberately: scope each agent to the minimum databases it needs, turn on line-by-line edit approval for anything client-facing, and control which models are allowed at Settings → Notion AI → General. For regulated or client-sensitive work, the Enterprise plan adds zero data retention with LLM providers and audit-log visibility over agent instruction and permission changes.

### Can Notion agents replace staff?

No, and the published results do not claim that. The best-documented case is James Lawley at Remote saving his team 20 hours per week on request routing — time returned to the same team, not headcount removed. Agents are strongest at triage and summarising, which is the work that sits around a job rather than the job itself.

### Why is my Notion agent ignoring a page I told it to read?

Almost always because the page was only linked in the instructions rather than added to the agent's access. Notion's documentation is explicit that linking to Notion pages in your instructions does not add them to the connected Tools and access. Open the agent's Settings, go to Tools and access, and add the page or database there.

### Do I need to know how to code to use Notion agents?

No. Custom Agents are configured in natural language — you can describe what you want and let Notion draft the instructions, triggers and access for you to review. Code only enters the picture with Notion Workers, which give agents deterministic tools to call, and Notion designed those to be built with AI coding agents rather than by hand.