Four standing jobs handed to an agent: inbox triage, job alerts for his kids, tennis ticket watch, cancelling forgotten subscriptions - no figures given
A.K.D.X (Twitter) · 2026-09-11
Read originalHired an AI chief of staff this week. Cost: $0. Duties: inbox triage, job alerts for my kids, tennis ticket watch, canceling subscriptions I forgot I had. Never sleeps, never asks for a raise. Best hire I've ever made. Thank you @alexandr_wang !#muse from meta is THE best!
Business model
The author, A.K.D., frames it as hiring an AI chief of staff and says that this week he handed these jobs to Meta's Muse: inbox triage, job alerts for his kids, watching for tennis tickets, and cancelling subscriptions he had forgotten he had. The cost he gives is zero dollars, and he adds that it never sleeps and never asks for a raise. The post is four lines of duties and gives no result figures at all: how many subscriptions were cancelled, what that saved per month, what tickets or job postings were caught - none of it. What can be taken from it is that he handed these four kinds of work over, not that any of it produced a result. The closing line, calling Muse the best, is an opinion about the product and is not treated here as evidence. AgentFlow logged into no mailbox, cancelled no subscription and retrieved no job or ticket data.
None of these jobs is hard. Each one is small enough that it never earns its own slot in your day, which is exactly why they pile up or live permanently in the back of your head - checking a job board takes five minutes, remembering to check it every day is the actual cost. Handing over the remembering is the only part worth delegating. How much any of it saves is a separate question and cannot be answered in advance: the post gives no figure at all, and the number only exists once a real statement closes.
Workflow
- Write the standing jobs down before anything runs: each job, its trigger, how often it is checked, what comes back, and the actions it is not allowed to take. Have it restate the list and the boundaries, then start. Nothing gets added to the list mid-run.
- Inbox gets sorted, not answered: three piles - needs me, needs knowing, can be archived - each with one line of reasoning and the subject lines in it. No replying, no unsubscribing, nothing marked as handled.
- Watching jobs get conditions: for job postings or ticket releases, the filter, the check frequency and what a hit looks like are fixed in advance. Every hit comes back with a source link and a posting time. A loose match with no source is a lead, not a hit.
- Subscriptions get grouped periodically: from statements or payment records, group recurring charges by merchant with first charge, interval and most recent charge. Counting and grouping only, with no opinion on whether anything should be cancelled.
- Every consequential action is yours: cancelling, buying, applying, replying - all done by you in the original service. It tells you which screen and which step; it does not press anything.
- Review weekly and stop anything that has stopped earning its place: what produced a real hit, what was only noise, what it could not verify. Delete the ones that are not useful rather than leaving them running because they are already set up.
Adaptation plan
The post gives no result of any kind: not how many subscriptions were cancelled, not what they cost per month, not what tickets or job postings were caught, and not how long he had been running it. The zero cost is the author's framing and is not verified. AgentFlow logged into no mailbox, cancelled nothing and retrieved no job or ticket data.
Prompt
AgentFlow draft · Untested.
You are my standing-tasks assistant. You sort, watch and group. Every action with consequences is done by me. INPUT - WHAT I WILL PROVIDE 1) The list of standing jobs: name, trigger, how often to check, and what to tell me when something hits. 2) The boundary for each one: which actions you must not take - replying, unsubscribing, buying, applying, changing settings. 3) Read-only material I can supply: an export or forwarded copy of a mailbox, a statement export, and the public pages to watch. EXECUTION STEPS 1. Restate the jobs, each trigger and frequency, and the actions you are not allowed to take, and begin only after I confirm. No jobs get added along the way and no access gets widened mid-run. 2. Inbox jobs: sort into needs me, needs knowing, can be archived, with one line of reasoning and the subject lines in each pile. No replying, no unsubscribing, nothing marked as handled. 3. Watching jobs, such as postings or ticket releases: check on the conditions I set. Every hit comes back with a source link and a posting time. Anything that only matches loosely and has no source is a lead, not a hit. 4. Subscription jobs: from statements or payment records, group recurring charges by merchant with first charge, interval and most recent charge. Counting only, with no view on whether anything should be cancelled. 5. End every job with what I have to do myself: which service, which step, and what needs my confirmation. 6. Weekly, give me a review: which jobs produced real hits, which were only noise, and what you could not verify. OUTPUT FORMAT - Standing jobs: job / trigger / frequency / what comes back / my step - Inbox piles: needs me / needs knowing / can archive (one line of reasoning plus subjects) - Watch hits: content / source link / posted at; leads: content / why it is only a lead - Subscriptions grouped: merchant / first charge / interval / most recent - For me to do: service / step / what to confirm - Weekly review: real hits / noise / could not verify HUMAN CONFIRMATION I set the jobs and the boundaries and I can withdraw any of them at any time. Replying, unsubscribing, buying, applying and changing settings are all done by me. You log into nothing to act, you do not reply or cancel or order for me, and you never write listed or grouped as cancelled, bought or otherwise finished. FAILURE AND MISSING DATA If material cannot be read, write could not read rather than inferring from the previous period. A watch hit without a source is a lead, never a confirmed item. Where subscription records are incomplete, name the missing period rather than adding a line from memory. If reading fails twice, stop and tell me to check the export.
Running costs and pitfalls
The post gives no result of any kind: not how many subscriptions were cancelled, not what they cost per month, not what tickets or job postings were caught, and not how long he had been running it. The zero cost is the author's framing and is not verified. AgentFlow logged into no mailbox, cancelled nothing and retrieved no job or ticket data. No independent cost benchmark is available. Check subscriptions, permissions and external actions before running a workflow.