A 24-hour AI agent is a different kind of tool

Notion just rebuilt the core of its AI agent. The biggest change sounds simple: a task that used to stop after five minutes can now run for over 24 hours. But a 24-hour AI agent is a different kind of tool. Before, you could ask it to write a draft or summarize a meeting — fast, self-contained jobs. Now you can set it to tackle something that takes real time: pull data from multiple places, write code, hit a problem, fix it, and check the result. The list of tasks worth handing off just got much longer.

The five-minute limit wasn't a design choice — it was an engineering constraint. Each session ran as a single job on a server. When the job ended, the agent's memory ended with it. The new version runs on a dedicated worker that saves its progress as it goes. If a server restarts mid-task, the agent picks back up where it left off. The agent no longer forgets what it was doing, which makes longer and harder work realistic.

A 24-hour agent only makes sense if it doesn't cost a fortune. Notion tackled this with a smarter model router called Auto. Instead of sending every step to the most powerful AI model, Auto picks the right model for each sub-task — cheap models for simple steps, the best for hard ones. Notion says this cuts average costs by about 90% versus using top-tier models for everything. The same monthly budget now runs about ten tasks where it used to run one, making all-day agents a real option rather than an expensive experiment.

The practical question to ask now isn't 'can AI do this in five minutes?' It's 'does this work take hours, and does it follow predictable steps?' Research, writing, data processing, multi-step code — all of that is now in scope. Notion also added a feature called 'Review for me,' where a second AI model checks the agent's planned actions before it takes them, making high-stakes tasks safer to delegate. For anyone building or running a business, the automation gap just got a lot smaller.