An agent that works all day will be interrupted
Notion's agents can now work on one task for more than 24 hours. Long runs get interrupted, so plan how they pick up again without doing anything twice.
Abhijit Chatterjee Chattopadhyay

On 19 October, Notion said its agents "can now work on a single task for more than 24 hours, up from five minutes."
That is from Notion's post on its rebuilt agents (opens in a new tab). It is a big change in what you can hand to an agent. It also changes what can go wrong.
Long tasks get interrupted
A five-minute task usually finishes before anything breaks. A task that runs for a day is likely to meet a restart, a deployment, a provider outage or a rate limit along the way. The question is no longer whether it gets interrupted. It is what happens when it does.
Notion's answer is durable execution: "Every session runs on a dedicated service with a queue-backed worker loop, so work keeps going through restarts and hiccups." In plain terms, the work is kept outside the agent, so a restart picks it up again instead of losing it.
Picking up again is the risky part
Resuming is safe for reading and thinking. It is not automatically safe for actions. If the agent sent an email, issued a refund or updated a record just before a restart, a careless resume does it twice.
Before you let an agent run for hours, check three things:
- Every action that changes something can be repeated safely. Give each one an ID, and have the receiving system ignore a second request with the same ID.
- Progress is saved after each step, not only at the end. Save what was done and what comes next, not the whole conversation.
- A person can see where it is. A short status the team can read: done, in progress, waiting, failed.
Long runs need limits too
A task that runs for a day can also spend for a day. Give every task a time limit, a step limit and a spend limit, and have it stop at whichever comes first. When it hits one, it stops and reports. It does not keep trying.
Notion also added a mode it calls "Review for me", which "uses a separate model to check whether an action is safe before the agent takes it." That is a sensible extra layer. It does not replace deciding in advance which actions an agent may take without a person, which I wrote about here.
Do this today
- List the actions your longest-running agent can take that change something: emails, payments, records.
- For each one, ask what would happen if it ran twice. Fix the ones where the answer is bad.
- Give every agent task a time, step and spend limit, and a status someone actually checks.
Your next step
If you want your agents ready to run for hours without surprises, book a free 30-minute 1:1 assessment. I look at your setup with you and tell you what the work would take.
Drafted with AI help, edited and approved by me.