What An Always-On Agent Mac Actually Does, Shift By Shift
An always-on Agent Mac works four shifts. Overnight, it runs the long jobs that take hours and nobody wants to babysit. Before dawn, it runs the scheduled sweeps that put a status report in your hands before you have made coffee. Through the business day, it sits idle and waits for events, a lead landing, a message arriving, a file changing, and responds in seconds. And once a week it runs the review that compounds what it has learned. None of those shifts can be covered by a machine that closes its lid, which is the whole argument for a dedicated box.
The night shift: long jobs nobody wants to watch
The most valuable hours on the Agent Mac are the ones when you are asleep. That is when the long work runs: the research job that reads a hundred pages and writes a brief, the content batch that drafts a week of posts from your transcripts, the data pull that reconciles last month's numbers across three systems, the rebuild of a site that takes forty minutes of agent time.
These jobs have two properties that make them wrong for a laptop. They take long enough that something will interrupt them if the machine is also being used for anything else, and they are not urgent, so nobody is sitting there to restart them when they fail. On the Agent Mac they kick off from a timer, run to completion, write the result somewhere you will see it, and the machine goes back to idle. If one dies, launchd restarts it. If the power blinks, the machine comes back on its own.
The habit this builds is simple: anything that will take more than ten minutes of agent time gets scheduled for the night rather than run in a chat window during the day. Within a few weeks you stop waiting for agents at all.
The pre-dawn shift: scheduled sweeps
Between roughly 5am and 7am the timer jobs fire. This is the shift most owners feel first, because the output lands in your inbox or your Telegram before you are at your desk. The morning sweep is the canonical example: what came in overnight, what is due today, what changed in the accounts you watch, what needs a decision from you. A second common one is the daily numbers: yesterday's leads, spend, revenue, pulled from the sources and written up in the same format every day.
The point of a scheduled sweep is not that an agent can read your inbox. It is that it reads it at the same time every day whether you remember to ask or not, and the format never drifts. That consistency is what makes it trustworthy, and trust is what lets you stop checking the underlying systems yourself.
On the Mac these are launchd timer jobs, each pointing at a prompt and a destination. The guide has a rake on this one: a scheduled job configured with a calendar interval that never fired, because the machine was asleep at the moment it was supposed to run. That bug is impossible on a correctly configured Agent Mac, and it is the whole reason the power settings come first in the setup.
The day shift: waiting for events
During business hours the Agent Mac is mostly idle, and that is correct. Its job in the day is to be awake when something happens. A lead fills out a form and the agent qualifies it and sends the first reply within a minute. A client texts and the agent drafts a response for a human to approve. A new call recording appears and the agent pulls the transcript and writes the follow-up notes. A file lands in a folder and the agent processes it.
Each of those is a trigger, not a schedule. Something external happens and the agent responds. This is the shift where the Mac's access to Apple apps earns its keep, because many of the events that matter in a small business arrive as iMessages, calendar changes, or emails in Mail, and a Mac is the only machine where an agent can see them. The bridge that connects the agents to Telegram, so you can talk to them from your phone, is also running in this shift, listening for you.
What "idle" looks like on the reference machine: the display asleep, the fans off, a handful of processes waiting, and 2.6 TB of disk untouched. Silent in the corner. The machine is doing its job precisely by doing nothing until it is needed.
The weekly shift: review and memory
The fourth shift runs once a week, usually over the weekend. This is where the agents look back: what ran, what failed, what the week's outputs had in common, what should change in next week's prompts. It is also where memory compounds. The agent that writes in your voice gets another week of your transcripts. The agent that qualifies leads gets another week of which ones turned into calls. The weekly job updates the files those agents read from, so next week's work is a little better than this week's.
This shift is the one owners skip, because it does not produce anything that looks urgent. It is also the one that separates a fleet that improves from a fleet that plateaus. Put it on the timer and forget it.
Why this does not fit on a laptop
Read back over the four shifts and count how many require the machine to be awake at a time you are not at it. All four. Overnight jobs need the lid open at 2am. The morning sweep needs the machine on at 5:30. Triggers need it listening at 11pm on a Saturday. The weekly review needs it running on Sunday.
A laptop fails every one of those by design. A desktop you also use for work fails the moment a software update reboots it mid-job. A cloud server passes the "awake" test but fails the "can see my iMessages" test and adds a monthly bill. The dedicated Mac is the only option that passes all of it, and the Agent Mac vs a cloud server works through that trade in detail.
If you are new to the idea entirely, start with what the Agent Mac is and why run agents on one. If you are ready to build the machine, how to set up an always-on Agent Mac walks through the order of operations and where people get stuck.
FAQ
How many agents can one Agent Mac run at once?
More than you will need at first. The reference machine runs a pod of agents concurrently alongside Redis and the messaging bridge, with most of its 72 GB of RAM free. The practical limit is usually how many parallel jobs you can keep track of, not the hardware. Start with three or four scheduled operations and grow from there.
What happens if the power goes out overnight?
The machine is set to restart automatically when power returns. With FileVault on, it lands at the disk unlock screen, which is the one step that cannot be fully automated. The guide covers how to handle that so an outage costs you a login, not a lost day. A small UPS removes most of the problem.
How do I see what the agents did while I was asleep?
Each scheduled job writes its output somewhere you already look: a message in Telegram, an email, a file in a shared folder. The morning status sweep is itself one of those jobs. You should never have to open the Mac to find out what happened. If you do, the operation is not finished being set up.