How does this update treat "Scheduled Tasks Claude creates for you" differently from "ones you manually set up yourself"?
This update's scope is explicitly limited to scheduled tasks "Claude creates for you" — meaning the kind you ask Claude in natural language to plan and set up, like "summarize last week's customer service emails for me every Monday." For this category, the default approval mode shifts to automatic approval, where the organization allows it. By contrast, if you manually configured a task step by step through the scheduling interface yourself, there's no indication this update touches its default approval mode at all. This distinction means that, starting with this update, a task Claude planned on your behalf and a task you set up with your own hands now sit at two different trust levels in the system's eyes.
Why would Cowork choose to shift the default toward automatic approval at this point, rather than keeping manual approval as the default?
A reasonable guess is that this is about usage friction. If every run of a Scheduled Task Claude built for you popped up an approval prompt first, that directly contradicts the core selling point of scheduling — that you don't need to be present to get it done. You'd still have to respond to the prompt the moment the task runs, undercutting the convenience scheduling is supposed to offer. By replacing "ask every time" with "the system judges for itself whether something looks unsafe," Cowork is essentially shifting part of the judgment burden from the user onto the system's real-time safety checks, bringing Scheduled Tasks closer to the fully automatic experience they were always meant to be — at the cost of placing more trust in the system's judgment.
Mechanically, how does the "pauses only when something looks unsafe" judgment actually work?
The official changelog doesn't disclose further detail on the specific rules or thresholds behind this judgment — it only offers the general description "looks unsafe," which means users currently have no precise way to know which kinds of actions trigger a pause and which don't. A reasonable guess is that this draws on the same safety system as Cowork's existing content classifiers — the system used to scan for Prompt Injection attempts and recognize malicious instructions is likely also what underlies the "should this pause" judgment here. But because there's no published, explicit list, users can't adjust or predict this judgment's sensitivity the way they can configure an approval mode directly — they can only treat it as an opaque last line of defense, not a rule they can actively configure.
For someone who's already had Claude set up Scheduled Tasks for them, what should they do now that this update is live?
The first step is an inventory: list every Scheduled Task Claude has created on your behalf and check, one by one, which approval mode each is actually using now — this update is a "default change," which doesn't necessarily mean every existing task gets forcibly migrated to the new rule, but it also can't be assumed they all stayed as-is; the safest move is to check the settings page directly. The second step is classification: sort these tasks by the nature of their actions into "fundamentally reversible" (organizing, summarizing, reading) versus "fundamentally irreversible" (sending, deleting, purchasing). For the latter category, even if the system's default has shifted to automatic approval, it's worth manually setting it back to manual approval, or at least "ask when needed" — don't let an action that can't be undone be gatekept entirely by an opaque automatic judgment system.
Cowork's October 1 release, v2.19675.0, quietly included a clause that's easy to skim past but actually changes default behavior: Scheduled Tasks that Claude creates for you now default to "Automatically approve," where your organization allows it.
The official changelog wording is direct: "Changed scheduled tasks that Claude creates for you to use 'Automatically approve' by default where your organization allows it, so their runs use tools without asking first and pause only when something looks unsafe." The key phrase is "that Claude creates for you" — this change targets scheduled tasks you asked Claude to plan and set up on your behalf, not ones you configured manually from scratch yourself. Whether automatic approval actually applies still depends on whether the organization-level policy permits it; individual accounts or organizations where the policy isn't enabled are unaffected.
The same release also fixed a related Edge Case: under "Skip all approvals" mode, if a connector tool was set by organization policy to "Restrict to Ask," that restriction previously wasn't being correctly enforced. It's now fixed so that "even in skip-all-approvals mode, newly started tasks will still ask per organization policy." This means an organization-level restriction on a specific tool can now more precisely override a user's overall approval-mode setting.
A Scheduled Task's core characteristic has always been that it runs unsupervised, with the user not present — something we've covered before when discussing scheduled task safety, and something the official guidance itself explicitly warns against using for anything involving sensitive files, sending messages, or making purchases. This update doesn't change that underlying "unsupervised" nature of scheduled tasks at all — it changes whether the system asks you first by default.
In other words, if you asked Claude to plan a scheduled task for you, it may previously have surfaced an approval prompt on each run, giving you a chance to glance at it before a tool was actually called. After this update, as long as the organization allows it, that prompt no longer appears by default — the task simply runs through, and only pauses if the system itself judges something to "look unsafe." That means the line of defense of "whether there's an approval prompt" has partly shifted from a setting the user directly controls to an outcome the system determines on its own.
If you're in the habit of asking Claude to plan and set up scheduled tasks for you, it's worth going back after this update to check which approval mode those tasks are actually using now — especially ones involving outbound communication, file deletion, or any irreversible action. For organization administrators, what needs checking is the current state of the "organization allows automatic approval" toggle, since flipping that switch affects every scheduled task Claude creates across the entire organization, not just one user's individual choice. The more reliable approach is to manually set any scheduled task that genuinely needs human oversight back to "manual approval" or "ask when needed," rather than assuming the system's own safety judgment will always catch everything it should.