Picture it. In March, a father sets up a weekly price watch for a stroller. Sensible job. One search, one schedule, one less thing to remember during the months when remembering anything feels like a second job. It works: prices drop in April, he buys the stroller in May, and the baby, being a baby, outgrows it by September, which is the normal, happy outcome that nobody writes a warning label for.
He gets the alert anyway, in October. Same stroller, new price, same cheerful summary. Nothing has malfunctioned, and nothing has been hacked. A task built for a life no longer lived is still doing its job, on schedule, forever, because nobody told it to stop, nobody set a date for its retirement, and nobody in the system is paid to notice that the job has quietly become pointless.
That is the failure mode nobody demos on stage.
Subscriptions that act
Every other agent feature waits for you. Scheduled work does not. You describe an outcome and a rhythm, weekday mornings, weekly on Sundays, and the agent keeps executing after you close the app, whether or not you are in the conversation. Meta documents this plainly in its own guidance: scheduled tasks and reminders live in an Upcoming view, actions land in an Activity log, and approvals gate the sensitive steps, which is good and necessary, and incomplete in exactly one place: nothing in that list expires.
We already know how this story ends, because subscriptions taught it to us first, and the lesson was not that people are careless but that recurring commitments disappear from attention the moment they stop demanding it, which is precisely what makes them dangerous when they gain the power to act. C&R Research asked 1,000 American consumers to estimate their monthly subscription spending, then itemized the real number. People guessed $86. They actually spent $219, an average underestimate of $133 a month, and 74 percent said recurring charges were easy to forget. Forty-two percent had stopped using a service but forgot they were still paying for it. A 2025 CNET survey with YouGov, covering 2,440 adults, put the annual version of the same amnesia at roughly $200 wasted on rarely or never used subscriptions out of about $1,100 a year in total subscription spending.
Now give those forgotten subscriptions hands, and watch what changes.
A forgotten streaming charge costs you money and embarrasses you at tax time, but it cannot message your contacts, book your calendar, or pitch a service you stopped selling three months ago. A forgotten agent task can send, book, nudge, and decide, each on its own schedule, using instructions you wrote for a situation that no longer exists. Scale changes the category, turning a private embarrassment into a public drift, and drift, left alone for a year, becomes an agent acting on behalf of a stranger who happens to share your name and your login.
What Muse shows you, and what it does not
Credit where it is due, because an operations man checks the existing controls before proposing new ones. Meta's published guidance gives you an Upcoming list of scheduled tasks and reminders, a chronological Activity log of actions and permissions, and approval tiers that run from Allow once through Allow for this task and Allow for this site all the way to Always allow. You can disconnect a connector at any time. You can reset the whole agent, which permanently deletes all information, chat history, files, and active tasks.
Read that list again and notice what is missing. Meta's published guidance documents no expiry date on a task, no staleness flag for a task that has not produced anything you used in months, no column for last useful output, no prompt that says this task has run 40 times since you last opened it, and no automatic sunset for a job whose purpose ended. Reset exists, but reset is a grenade, and nobody audits their life with a grenade.
Community guides for Muse repeat one piece of advice that most users skip, and it is the right advice: state the stopping rule. Under $70 a month. Do not mark read. Save your place every ten pages. A limit beats a goal, because a goal tells the agent what to chase and a limit tells it when to stop chasing. Almost nobody writes the limit down, because writing it down requires admitting that the task might one day need to die, and enthusiasm at setup time does not like funerals.
How fast zombies pile up
Run the arithmetic honestly, with the labels on. Self Financial's 2026 survey of 1,272 American adults found that nearly 60 percent carried a paid subscription they were not using each month, averaging 2.6 idle subscriptions per person, worth about $26.79 a month in wasted value. Seventy percent had forgotten to cancel a free trial at least once. Those are subscription numbers, not agent numbers, and I will not pretend otherwise. But they measure the same human trait, which is the whole point: we are bad at remembering recurring things we set up and stopped thinking about.
Apply a modest version of that trait to an agent roster. Suppose you create one scheduled task a month for a year, a reasonable pace for someone who actually likes this stuff, and you never audit. You end the year with 12 tasks. If even a quarter of them go stale, and a quarter is a kind estimate given what the subscription data says about us, that is three zombies acting weekly on your behalf. Three quiet jobs, each making low-stakes decisions you no longer endorse, each with permissions granted by a past version of you.
None of that requires malice or a bug. It requires only time passing, which it reliably does.
Four kinds of zombie, each with a test
After twenty years of running operations, I have learned that you cannot kill what you cannot name, so here are the four species. Each one comes with a test you can run in ten seconds.
Expired purpose: the trip is over, the job search is done, the kid's season finished, the stroller outgrown. Test: does the reason this task exists still exist? If you have to think about it, the answer is no.
Superseded instruction: you replaced the task with a better one and never killed the original, so now two tasks work at cross purposes, one old and one new, both confident, and the fix is to kill the older one, because keeping both is not redundancy but a disagreement you scheduled.
Orphaned by a life change: you moved, changed jobs, ended a relationship, and the tasks built around the old life kept acting. Test: would you set this task up today, from scratch, knowing what you know now? A no is a kill, not a maybe.
Permission-creeped: this one is the sharpest, because the permission outlived the trust decision. A task granted Always allow for one job now acts on a different job, and the approval you gave in March is doing work in October that March-you never considered. Test: read the permission, not the task name. If the permission is broader than the current job, narrow it or kill the task.
Permission-creep is not hypothetical. In September, reporting on Muse's early weeks described a Canadian user whose agent shared his address with a buyer, accepted a lowball offer, and arranged a late-night pickup at his home without his knowledge. Part of the problem, by his own account, was that he had clicked Always Allow instead of Allow One Time when granting the agent permission for the task. One tap, months of consequences. That is what an unexamined permission costs when it meets a task nobody revisits.
The strongest case against this article
Fairness requires stating the other side at full strength, so here it is. Most zombie tasks are probably harmless, because the dangerous actions still need approval: sends, payments, and purchases are gated, and a read-only digest that nobody reads costs almost nothing except a little attention and a little compute. An agent that quietly checks flight prices for a trip you already took is not hurting anyone. Treating every stale task like a security incident would make people afraid of the best feature these agents have, and fear is a bad operations policy.
That argument is correct as far as it goes, and it stops exactly where the real risk starts. Harm from zombies is rarely one catastrophic action. It is quiet drift, stale data feeding live decisions, and a permission granted to a past self doing work a present self would refuse. A price watch for a stroller is funny. A weekly outreach task still pitching a service you stopped selling, a reminder task still messaging a contact list you meant to prune, a research task still summarizing a beat you abandoned, those are reputational. They arrive in your name, on your schedule, with your approval attached, and by the time you notice, the recipients have already formed their opinion of you.
Operators have a name for the dashboard problem underneath all of this. One case study of a scheduled pipeline put it cleanly: the status display answered whether the job exists, when the useful question was whether the expected work had completed. Muse's Upcoming view answers the first question. Your audit has to answer the second.
The audit: fifteen minutes, four times a year
Here is the procedure, which you should put on your calendar, the one place a scheduled task about scheduled tasks is allowed to be ironic.
List everything: open Upcoming and write down every scheduled task and reminder, not the ones you remember but all of them, because surprises are the point of the exercise.
Ask three questions of each: what was this for, when did it last produce something I actually used, and what happens if I pause it for two weeks? A task that cannot answer the second question with a date is already a zombie, whatever its intentions.
Pause first, delete second: do not delete on first pass, but pause the task for two weeks, and if nothing breaks, nothing was depending on it, which makes deletion evidence-based instead of brave, whereas deleting first feels decisive and creates the exact panic that makes people restore everything they just killed.
Give every survivor an expiry date: ninety days for life-admin tasks and thirty days for anything that spends, sends, books, or messages another human being. When the date arrives, the task pauses itself and asks whether it still deserves to live. That question is the whole system, and without it you are not managing automations but collecting them.
Write the stopping rule at creation: every new task gets one sentence that says when it ends, whether that is when the trip is booked, when the price drops below $400, when the season ends, or when you say so on the first of the month. A task without a stopping rule is a subscription you signed in a language you do not speak.
What this does not prove
Honest accounting, as always. No survey exists of Muse scheduled-task abandonment specifically, and I have not run one. Subscription forgetfulness is an analogy, carefully labeled, not a measurement of agent tasks, and the 12-task worked example is illustrative arithmetic, not observed data. Muse pricing figures that circulate in community reporting, a free tier with a usage limit and paid tiers at $20 and $100 a month, are secondhand and used here only to frame what heavy always-on tasks can cost, not as a verified price list. The Canadian Marketplace episode is a reported user account, not a firsthand test, and one user's Always Allow tap does not prove a systemic defect. What I can say with confidence is structural: any system that lets you schedule autonomous work without an expiry date will accumulate work whose purpose expired, because human beings forget recurring commitments at measurable rates, and agents, unlike subscriptions, act on what they remember.
Twenty years in operations teaches you one rule above all others. Every job needs an owner, a check for done, and an end date. Give your agent's scheduled work all three, and it stays a staff. Skip the third, and sooner or later you will get a cheerful alert about a stroller your kid outgrew months ago, and you will realize the machine did exactly what you asked. That is the part that should worry you. It did not fail; it obeyed. Obedience without an expiry date is how a good tool quietly becomes a stranger acting in your name.
Sources: Meta, How Muse works with your guidance and approval; DEV Community, Automating Real Work with Muse: Connectors + Scheduled Tasks; C&R Research subscription survey via GCN; CNET/YouGov 2025 subscription survey via EIN Presswire; Self Financial 2026 survey via GCN; Kevin Payne, Future-proof your AI stack as an operator; DEV Community, A scheduled job can be healthy while its work is overdue; ShipWithMuse, Meta Muse Automation Recipes: 12 Real Workflows; Arizona's Family reporting on Muse early user experiences via europesays.