Always-on agents are having a moment. These are AI agents that keep working for you in the background, for days or weeks, without anyone watching. Anthropic has just published a practical guide to building them, and its first lesson is a pleasingly honest one: “An always-on agent is not actually ‘always on.'” Like a cat, it sleeps most of the day and wakes the moment something lands in its bowl.
If you juggle several calendars, this matters more than it might seem. The agents that will soon book, move and decline meetings for you are exactly this kind of software. Whether they’re any good won’t come down to the model alone. It comes down to the plumbing underneath.
What Anthropic says always-on agents need
In Building always-on agents, published on 13 November, Anthropic’s William Hu argues that these agents run the same core loop as any other agent. What’s different is the infrastructure around them, built from five primitives:
- An event inbox. Webhooks, scheduled jobs and messages are written to a durable list rather than calling the agent directly, so events aren’t lost and runs don’t compete. Events are deduplicated, and “an inbox should only have one active worker at a time.”
- Timers. Follow-ups and recurring schedules live in their own store, and a scheduler turns them into inbox events when they’re due.
- Memory. State lives in durable storage, and it “should only hold what the agent can’t look up again.”
- Permissions. The agent acts with your access while nobody is watching, which the guide calls the area of greatest risk. Credentials stay outside the agent, and important actions still go to a person.
- An outbox. Messages to the user are paced, and the agent should “start quiet and then earn more of the user’s attention over time.”
The guide puts the renewed interest down to models with better judgement and to intelligence getting cheaper, which makes long-running work affordable. None of the five primitives is new, though. They are the bones of any system that has to react to the outside world reliably.
We’ve been running on the same bones
Reading the guide felt familiar. Calendrz isn’t an AI agent, but availability mirroring has the same job description: wake up when one of your calendars changes, work out what that means for all the others, act, and go back to sleep. Here’s how the primitives map onto what we built.
The inbox: never act on every ping
Google Calendar and Microsoft 365 notify us when a calendar changes. Those notifications arrive in bursts: move one meeting and several calendars can ping at once. We don’t run on each one. We collect them in a buffer, skip the ones already there, and every 30 seconds turn the batch into a single run per Profile. It’s the guide’s inbox and batching advice, arrived at by watching our own traffic.
One worker per inbox
Only one task runs for a Profile at any time. If the same kind of task is already running or waiting, the new request is dropped; anything else queues behind it. This sounds fussy until you imagine two runs racing over the same meeting and each creating a block for it. A calendar tool that double books you by itself would be a fine piece of irony.
Timers, and a lesson we learnt
Notifications are the fast path, not the only one. A regular check every few minutes catches anything a notification missed, and other timers refresh the access for connected accounts and renew the notification channels with Google and Microsoft, which expire. Our founder, Liviu Tudor, found that we used to renew a channel only once it had already expired, leaving a window of up to half an hour where changes went unheard. The fix was to renew 30 minutes early. Timers are easy to add and easy to get subtly wrong, so build in slack.
Memory as pointers
Every block that auto-blocking creates carries a small pointer back to the event that caused it. The next run uses it to find the block, move it when the original is rescheduled, or remove it when the original is cancelled. That is the guide’s advice in miniature: keep a pointer to the source, and look the rest up fresh.
A quiet outbox
Once your connected accounts are set up, Calendrz mostly stays out of your way. We email you when something needs your attention, such as a connected account that has lost its access, and not to tell you that everything is fine. Start quiet, earn attention.
What to take from it
When you weigh up an AI assistant that will act on your calendars, ask the boring questions. What happens when two changes land at once? What does it do with a notification it missed? Where do its credentials live, and which actions does it check with you first? The answers tell you more than any demo.
We also think the best software of the next few years is software you rarely open. Agents will do the clicking for you, and the dashboard becomes something you visit when something is wrong. Always-on, done well, looks a lot like nothing happening, which is exactly how a calendar with no double bookings should feel.
If you’re juggling several calendars and want them to stay in step without you babysitting them, try Calendrz for free. Set it up once, and let the plumbing do the rest.







