How to build a Custom Agent

TL;DR

Start by identifying repetitive, manual work that consumes time each week and can be standardized for the whole team.

Briefing Cornell Notes

Briefing

Custom Agents in Notion are built to take over repetitive, low-impact-but-time-consuming work—so teams can stop routing the same questions and start using their time on higher-value tasks. The core move is identifying a recurring workflow (often weekly) that’s manual, routine, and easy to standardize, then deciding whether setting it up once would benefit the whole team.

A practical example centers on a Q&A agent for a customer support Slack channel. The weekly problem is familiar: the same billing and process questions keep coming up—like where to log billing-related issues or how to handle edge cases. Instead of answering from scratch each time, the agent is configured to search the team’s help center and internal knowledge base, then respond with answers drawn from those sources. The setup begins in the workspace sidebar under the Agents section, where a user can build from scratch, start with a template, or—recommended for first-timers—chat with a Custom Agent to describe the desired outcome.

Once the goal is clear, the build process focuses on three pillars: triggers, instructions, and connected sources. Triggers determine when the agent runs, either event-driven (such as when the agent is @mentioned in Notion or when a Slack message arrives) or scheduled (like running every Monday morning or at the end of the day). Instructions define what “good” looks like—using clear outcomes, example responses, and the right references. The agent can be guided with simple if-then logic, such as: if the answer exists in the help center, reply with a short summary and link; if not, say it doesn’t know and route the question to the appropriate team.

The agent also needs access to the right tools and pages. Even when pages and tools are pre-populated, permissions must be reviewed. In most cases, view access is enough; edit access should be granted only when necessary. For Slack specifically, the configuration controls which channels the agent can read from or post to—private DMs and unconfigured channels remain off-limits.

Before wider rollout, testing matters. A quick check in the target Slack channel helps confirm the tone and verify that the agent is pulling from the correct pages. If something feels off—like wording that’s too cold—the instructions can be adjusted and the agent iterated on. Operational visibility comes through the Activity tab, which logs each run and whether it behaved as expected. Notion also provides Version History so changes can be rolled back.

Finally, model behavior can be left on “auto” or set to a specific model for more control. After the agent works in Slack, the Sharing panel can extend access to teammates with permissions ranging from “can view and interact” to “can edit” or “full,” depending on whether they need to update settings. The result is a system that continuously handles small repetitive requests in the background, freeing people for work that actually requires human judgment.

Cornell Notes

Notion Custom Agents are designed to automate repetitive, low-impact tasks—especially recurring Q&A—so teams can spend time on higher-value work. The process starts by spotting a workflow that’s routine and manual, then defining a clear outcome for the agent. Triggers decide when the agent runs (event-driven like Slack messages or scheduled runs), while instructions specify how to answer using the right sources and if-then rules. Connected tools and pages require careful permission checks, typically granting view access and limiting edit access. After testing in the target channel, teams can iterate using the Activity tab and Version History, then share the agent with teammates using appropriate permissions.

How do Custom Agents decide when to run, and what are the main trigger types mentioned?

Triggers control the timing of an agent’s actions. They can be event-driven—running when something happens, such as a new Slack message or a page added to a database—or scheduled—running automatically on a defined timetable like every Monday morning or at the end of the day. In the Q&A example, the agent is set to run when it’s @mentioned in Notion and when a message is sent in Slack.

What makes “instructions” effective when building a Q&A agent?

Instructions work best when they clearly define the goal and the expected outcome, like a capable teammate receiving a mission rather than a step-by-step script. The build process emphasizes showing what good looks like with examples, pointing the agent to the right sources (help center pages, internal knowledge base, or shared docs), and using simple if-then guidance for judgment. Example: if the answer exists in the help center, return a short summary plus a link; if it doesn’t, admit it doesn’t know and route the question to the right team.

Why do permissions matter even if the agent already has pages and tools populated?

Permissions determine what the agent can access and where it can operate. Even when pages and tools are pre-populated, the agent’s access must be reviewed and confirmed. The guidance is to use view access in most cases and grant edit access only when truly needed. For Slack, the configuration specifies exactly which channels the agent can read from or post to; it won’t see private DMs and won’t respond in channels that weren’t configured.

What’s the recommended approach for testing and improving an agent before broad rollout?

Testing should happen in the real target environment. The guidance is to go to the Slack channel, post a question, and check whether the agent pulls from the correct pages and whether the tone feels right. If the responses need adjustment—such as sounding warmer—the instructions can be edited to add examples, clarify formatting, or @mention different pages. Iteration is supported by the Activity tab (showing each run and results) and Version History (allowing confident experimentation and rollback).

How can teams control the agent’s model behavior and who can use it?

Model selection can be left on “auto,” which lets Notion pick the best model for each request, or a specific model can be chosen for more control. For team access, the Sharing panel controls permissions: teammates can be granted “can view and interact” to chat with the agent, or higher permissions like “can edit” and “full” if they need to update the agent’s settings.

Review Questions

What trigger types are available for Custom Agents, and how would you choose between event-driven and scheduled triggers for a support workflow?
Describe an if-then instruction pattern you could use to ensure a Q&A agent responds with citations when it finds an answer and routes questions when it doesn’t.
What permission checks should be performed for connected tools like Slack, and why is view access often preferred over edit access?

Key Points

1
Start by identifying repetitive, manual work that consumes time each week and can be standardized for the whole team.
2
Define a clear outcome for the agent in chat, then rely on well-written instructions rather than spelling out every step.
3
Use triggers to control when the agent runs, including event-driven triggers (e.g., Slack messages) and scheduled triggers (e.g., weekly runs).
4
Guide the agent’s decision-making with simple if-then rules and by pointing it to the correct knowledge sources.
5
Review tool and page permissions carefully—typically grant view access and restrict edit access to what’s necessary.
6
Test the agent in the target Slack channel to validate both answer quality and tone before expanding access.
7
Iterate using the Activity tab and Version History, and share the agent with teammates using the right permission level.

Highlights

Custom Agents are positioned as a way to remove weekly, repetitive work—like answering the same support questions—so people can focus on tasks that need human judgment.

Effective Q&A instructions combine source targeting (help center and internal knowledge base) with if-then logic for when answers exist versus when they don’t.

Even with pre-populated pages and tools, permission checks are essential; Slack access is limited to configured channels and excludes private DMs.

Activity logs and Version History make it practical to test, refine tone and formatting, and roll back changes if needed.

Notion’s “auto” model setting can be left on for convenience, while the Sharing panel controls who can interact versus who can edit settings.

Topics

Custom Agents
Q&A Automation
Triggers and Scheduling
Permissions
Iteration and Versioning