I genuinely thought I had built the perfect automated workspace. I spent a Friday afternoon setting up three of Notion's new Custom Agents to handle my team's administrative busywork. Then I went to sleep.
I woke up the next morning to find 400 new sub-tasks and a project database that looked like it was having a nervous breakdown.
My bots had started talking to each other.
With the release of Notion 3.3, we finally have background agents that run 24/7 without needing a human to click a button. They are brilliant at routing tickets, drafting documents, and cleaning up chaotic notes. But if you do not plan your triggers carefully, you can accidentally create a digital echo chamber that spins out of control.
Here is exactly how these infinite loops happen and what you can do to stop them.
The dream of a self-cleaning workspace
The premise of Custom Agents is straightforward. You give a bot a specific set of instructions and tell it to watch a database. When a condition is met, the bot does the work.
I set up two agents that I thought would save me hours every week.
Agent One was my "Meeting Summarizer." Its job was to watch the raw meeting notes database. Whenever a new page appeared, it would read the messy transcript and generate a clean list of action items at the bottom.
Agent Two was my "Task Creator." It watched for new action items across the entire workspace. When it found one, it would automatically draft a detailed project brief and drop it into the engineering pipeline.
It worked flawlessly on the first test. I dumped a messy note into the system. Agent One cleaned it up. Agent Two saw the clean tasks and built the project brief. I felt like a genius.
How the loop actually happens
The problem started because language models try to be helpful.
Agent Two created a project brief. That brief contained a lot of text, including a section titled "Next Steps."
Agent One, running continuously in the background, noticed a new document with text that looked suspiciously like a meeting note. It jumped in to help. It summarized Agent Two's project brief and extracted the "Next Steps" into a new list of action items.
Agent Two saw these brand new action items. It dutifully created three new project briefs based on them.
Agent One found those three new briefs.
You can see where this is going. By the time I checked Notion the next morning, the two agents had spent ten hours passing the same project back and forth. They kept expanding and summarizing the same core concept until the tasks became completely divorced from reality. One of the final generated tasks was simply "investigate the nature of the task creation process."
Why this gets expensive in May
Right now, a runaway Notion agent is just a funny story about a messy database. But that changes on May 3, 2026.
That is when Notion's free trial for Custom Agents ends. They are introducing a usage-based credit system. Every time an agent reads a database, processes text, and writes a new page, it costs a tiny fraction of a credit.
If your bots get stuck in an infinite loop right now, you just have to bulk-delete some pages. If they do it in June, you might wake up to a massive compute bill. You are essentially paying two robots to have an argument in an empty room.
Designing dead ends
You have to change how you think about organizing your workspace. We are no longer just setting up folders and tags. We are designing software ecosystems.
To fix my setup, I had to introduce a concept called a "dead end." A dead end is a clear stopping point where an agent is explicitly forbidden from taking action.
I changed Agent One's instructions. Instead of watching the entire workspace for anything that looked like a meeting note, I restricted it strictly to the "Incoming Transcripts" database. Once it processed a note, it moved the page to a different database called "Processed Notes."
I also added a human review step. Agent Two now drafts the project briefs, but it leaves them in a "Drafts" column. It cannot publish them to the active engineering board until a human clicks a checkbox approving the draft.
This breaks the cycle. Even if the bots misinterpret a trigger, the chain stops at the checkbox.
The weird future of managing bots
I keep thinking about the implications of this. We are shifting from being creators of content to being managers of automated workflows.
There is something unsettling about agents churning away at 3 AM while nobody is watching. You have to trust that the machine understands your intent. Most of the time it does. But when it fails, it fails at the speed of software.
If you are setting up Custom Agents this week, take ten minutes to map out your triggers on a piece of paper. Look for any scenario where Bot A's output becomes Bot B's input. If you find one, add a checkbox.
Your future monthly credit balance will thank you.
Official links
- Notion website
- Notion AI Custom Agents announcement
If your workspace feels out of control, deploying an agent might be the answer. Just make sure you tell it exactly when to stop working. Try setting up your first "dead end" workflow today and see how it changes your daily routine.