How to Build a Custom AI Agent in Notion (2026)

On October 19, 2026, Notion overhauled its AI agents: a sixth-generation harness that lets a single agent task run for more than 24 hours (up from five minutes), a smarter "Auto" model router, and a new two-pool credit system. If you are on a Business or Enterprise plan, this update makes Notion's Custom Agents something you can genuinely rely on for recurring work. This guide walks you through building your first one step by step.

What's new in October 2026

It helps to know what changed first, because the new engine changes how you should design agents.

Notion rebuilt the infrastructure behind Notion Agent and Custom Agents. According to the company's official announcement, each session now runs on a dedicated service with a queue-backed worker loop, so work keeps going through restarts and hiccups — that is how a single task can now run for more than 24 hours. Agents also get their own computer and browser, letting them run short scripts as code instead of spending language model tokens on the same work.

The headline numbers are Notion's own claims: Auto delivered results similar to frontier models like Opus 5.5 and Astra at about 90% lower average cost across 20-plus common use cases; agent runs are up to 50% faster on web and mobile; and Custom Agent runs cost roughly 30% less in Notion's testing. A new "Review for me" mode uses a separate model to check whether an action is safe before the agent takes it — a genuinely useful guardrail for agents with write access.

The new credit system is live now for new customers, and existing customers move over on November 2, 2026. Business and Enterprise seats include 1,500 Standard credits per month (work where Notion picks the model) and 500 Select credits per month (work where you pick the model, or run your own Custom Agents). The pools are separate on purpose: a few complex tasks or frequently running agents could exhaust a shared allowance, so everyday Auto usage is protected from your hungry agents.

Notion screenshot

Before you start: requirements and cost

Custom Agents are not available on every plan. According to Notion's help documentation, you need a Business or Enterprise plan, and you must build, edit, and interact with agents on desktop or web — the creation UI is not available on mobile. Admins in your workspace also control who can create agents, so if the option is missing, check with your workspace admin first.

On pricing: your own Custom Agents use your Select credits, and how quickly they burn them depends on the complexity of the work and the model they use. Team Custom Agents use Select credits by default even when other people run them. The good news: you no longer need to buy credits before building Custom Agents — the included allowance gets you started, and Notion's own estimate is that the vast majority of users are unlikely to run out of monthly credits.

Step 1: Open the Agents section and choose a creation path

In your Notion workspace's left sidebar, look for the Agents section and click the + button. Notion offers three starting points:

Create with AI chat — describe what you want in natural language, and Notion AI drafts the instructions, triggers, and access settings. Review, iterate, save. This is the fastest path and the best default for a first agent.
Create from a template — start from a pre-built template and customize it. Templates are also worth browsing for examples of good instruction design.
Create from scratch — start with a blank instructions field, for when the AI draft keeps missing the mark.

For this guide, pick Create with AI chat and describe one concrete job. Good example: "Every Monday morning, go through my Projects database and give me a summary of what is due this week, what is overdue, and what is blocked." Bad example: "Help me with my tasks." The more specific the input, the better the draft.

Step 2: Write instructions that actually work

This is where Custom Agents are won or lost. Review the draft Notion AI generated and rewrite it until it states, in plain language: what the agent should do, how it should behave, and what work it handles — with examples where you have them.

A strong instruction set has four parts:

The trigger condition restated. "Run every Monday at 9:00 AM" or "When a new page is added to the Leads database."
The exact output format. "A bulleted summary grouped by Due this week / Overdue / Blocked, with a link to each page. Keep it under 300 words."
The scope boundaries. "Only read the Projects database. Do not edit pages or change properties — summarize only."
The escalation rule. "If more than 20 items are overdue, flag that at the top instead of listing every one."

Name the agent after the job it performs ("Weekly Project Digest"), not a generic concept ("Helper"). You will thank yourself when you have five agents running and need to read the activity log.

Step 3: Set triggers and a schedule

Custom Agents are designed to run automatically in the background based on triggers and schedules. You configure them once, and they keep going. Go to your agent's settings and choose how it runs:

Recurring schedule — daily, weekly, monthly, or custom patterns. The weekly digest above runs on a schedule.
Slack events — a message posted to a channel, an emoji reaction, a thread started, or the agent being mentioned. Classic use case: triage every message in a #bugs channel and file structured bug reports.
Notion events — a page added to a database, a page updated, a comment added, or the agent being @mentioned. Classic use case: when a new lead lands in the CRM database, create a task for the right owner.

You can combine multiple triggers on a single agent — for example, a standup agent that runs on a daily schedule and also when someone @mentions it in a comment asking for an on-demand update.

Step 4: Grant access — and only access

Custom Agents act only on the pages, databases, and external apps you explicitly grant them access to. They never get full workspace access by default. Linking a page inside the instructions does not grant access on its own — this is the step where most first-time setups go sideways. If your agent cannot find your data or posts to the wrong place, check the access settings first, not the instructions.

Grant the minimum the job needs: the database it reads, the database or page it writes to, and the Slack channels it watches — nothing else. If the agent needs to post to Slack or read your email and calendar, connect those sources deliberately.

Step 5: Test, review, and publish

Before you switch the real trigger on, run the agent manually and check the activity log: did it read the right data, is the output format usable, and did it stay in scope? If the format drifts, tighten the instructions with an example of the exact output you want.

Turn on "Review for me" for any agent that writes, edits, or sends. It adds a safety check by a separate model before the agent acts — cheap insurance for agents with write access.

Once the manual runs look right, publish. Published agents continue running in the background: schedules fire on time, and event triggers react as things happen. Use version history to review or restore past configurations when you iterate.

Keep it credit-smart: five tips for the new system

The October update changed the economics of agents, so design for them:

Let Auto pick the model unless you have a reason not to. Auto routes each task to the model that balances quality, speed, and cost — simple requests go to fast, efficient models; harder work goes to more capable ones. Notion's own testing puts the savings at about 90% versus always using a frontier model.
Remember the two pools. Everyday asking and writing uses Standard credits (Notion picks the model). Custom Agents always draw from Select credits — and you can start building with the included allowance without buying anything first.
Prefer schedules over frequent event triggers for heavy jobs. A digest that runs once a morning is predictable; an agent reacting to every message in a busy channel can surprise you. Scope access tightly — every extra database is more context processed on each run.
Watch the new visibility tools. Members get a clear view of their own usage, and admins will get the AI Command Center: usage by member, feature, and agent, alerts as the workspace nears its limit, and spending limits. Set limits before a shared team agent becomes popular, not after.

Limits worth knowing

A few honest boundaries so you set expectations correctly. ChatGPT subscription connections (Plus, Pro, or Go) work in your personal Notion Agent first, not Custom Agents — at least at launch. Bring Your Own Key is in private alpha for Enterprise customers, not generally available, and the AI Command Center is a preview, not yet available to admins. And while agents can now run for over 24 hours, long-running work should be broken into verifiable checkpoints where possible: an agent that compounds mistakes for a day is worse than one you review every few hours.

Summary

Notion's October 2026 update makes Custom Agents meaningfully more reliable: 24-hour task windows, a harness that survives restarts, agents with their own computer and browser, and a credit system where Auto stretches your allowance. The build process itself has not changed — create from the Agents section, describe the job, write tight instructions, set triggers, grant minimal access, test, and publish. Start with one boring, recurring, verifiable workflow, run it under the new Select credits, and let the activity log tell you whether the instructions are good enough.

参考来源

Notion, "Agents, rebuilt to do more for less" (official blog, October 19, 2026) — notion․com
Notion, "Inside Notion's upgraded model router, Auto" (official blog, October 20, 2026) — notion․com
Notion Help Center, "Custom Agents in Notion" (help documentation) — notion․com
WebProNews, "Notion Rebuilds Its AI Agents, Upgrades Auto Model Router and Moves to a New AI Credit System" (October 20, 2026) — webpronews․com
Technobezz, "Notion AI Adds Two Credit Pools and Longer Agent Tasks" (October 20, 2026) — technobezz․com