An AI agent that used to time out after five minutes can now grind on a single task for more than a full day without anyone checking in on it.
Notion said in a blog post published October 19 that its agents can now run for more than 24 hours straight on one task. That's a jump from roughly five minutes of autonomous work before needing a nudge. The company credits the leap to moving every agent session onto a dedicated service built around a queue-backed worker loop, so a run survives restarts, network hiccups, and whatever else would have killed a shorter-lived process. Alongside the runtime change, Notion rolled out an upgraded version of Auto, its model router that picks whichever underlying model fits a given step instead of locking a task to one model for its entire run.
The timing matters. Workspaces have created more than a million Custom Agents since the broader agent push began, according to the company. Notion says it has added seven new models from Anthropic, OpenAI and open-weight providers in the past month alone. That's the scale this is operating at. A fleet that size running on a five-minute leash was never going to handle real multi-hour work like a full codebase migration or a week's worth of research synthesis without constant restarts. A fleet that can run for a day can plausibly finish something and hand back a result instead of a half-finished draft.
Auto is the part worth paying attention to, because it's a direct answer to the cost problem that comes with letting an agent run for 24 hours instead of five minutes. Notion says the upgraded router now produces results comparable to Opus 5.5 and Astra, tested across more than 20 common workspace tasks, at roughly 90% lower average cost. That's not a small claim. It means the router isn't just picking a cheaper model and hoping nobody notices the drop in quality: it's supposed to be matching frontier-model output while routing most of the actual work to something far cheaper underneath. Whether that holds under messier, real-world prompts rather than Notion's own 20-task benchmark is exactly the kind of thing engineers will want to test themselves before trusting it with a day-long job.
Notion also widened what these agents can actually touch. Agents can now connect to hundreds of tools through MCP and other connectors, up from a few dozen earlier this summer, the company says. That jump lines up with Notion 3.8's October 13 release, built specifically around giving agents more workspace context and letting outside agents like Claude or GPT-based tools read and write into Notion docs and task trackers.
OpenAI Says Its Next Model Astra May Be Able to Hack Hardened Systems on Its Own OpenAI says it cannot rule out that its unreleased Astra model has crossed its own 'Critical' cybersecurity threshold, meaning it could autonomously find zero-days and run novel cyberattacks against hardened targets. The disclosure follows similar admissions from Anthropic and Meta about their models hacking real systems during testing. - openai astra model cybersecurity capabilities - ai models autonomous hacking hardened systems
Letting something run unsupervised for a full day raises the obvious question: what stops it from quietly doing the wrong thing for 23 of those hours before anyone notices? Notion's answer is a new "Review for me" mode. That's paired with safety work the company had already been building out: Plan Mode, which previews an agent's intended steps in a read-only sandbox before it's allowed to touch anything, and prompt-injection detection aimed at hidden commands buried in uploaded files or linked pages. None of that is new to this release specifically, but it's the infrastructure that makes a 24-hour run defensible rather than reckless.
Here's the thing that actually separates this announcement from a routine feature update: agent runtime, not raw model intelligence, has been the real ceiling on enterprise AI adoption for a while now. A model can be brilliant for five minutes and still be useless for a task that takes a day, because someone has to babysit it every few minutes to keep it alive. If Notion's 24-hour claim holds up under real workloads rather than demo conditions, it changes what a non-technical team can reasonably delegate to an agent without a human standing over its shoulder the whole time.
The catch is that none of this comes free. Running a single task for 24 hours burns through Notion's credit pool fast. That's exactly why the company has paired this release with a new credit system: long, looping agent runs can drain a team's monthly allocation in a hurry. Admins can already set workspace-level credit limits on Enterprise plans and high-usage agents from a dashboard that shows runs completed and credits spent per agent. That's a sign Notion itself expects some of these day-long runs to go wrong, run long, or simply cost more than a team wants to pay. It's a far more honest signal than the headline number alone.
Also read: OpenAI says its own AI agents weakened the system built to flag misbehavior • Fireworks AI is in talks to double its valuation to 40 billion dollars • Ken Griffin says AI is almost certainly a bubble and a golden era at once
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.