Always on

An autonomous agent can be always on: it keeps working in the background. It wakes on a timer and whenever something happens, and every time it wakes it carries on the same conversation it was having before, so it remembers what it did yesterday.

Always on is not the same as a schedule. A schedule runs one task at a set time (every weekday at 9:00) and posts the result. Always on is a standing job: it wakes often, looks at what changed, and picks up where it left off. An agent can have both.

Turning it on

Open the agent and find the Always on card (on the agent’s page and in the builder), then Set up always on. You choose:

• What it keeps doing. Standing instructions, read at every wake. Write them the way you would brief a colleague: “Keep the refund queue moving. Answer what you can, and tell me about anything over $500.”

• What wakes it.

  • A timer: every so many minutes or hours. Your plan sets the shortest timer (see Limits).
  • Messages on its channels: pick any of its Slack, email, Teams, webhook and other messaging channels. A webhook delivery wakes it with what was sent. On every other channel, the people writing keep their own chats exactly as before; the agent is only told that someone wrote, never what they said, so one person’s words never turn up in an answer to someone else. With a timer on, that note waits for the next wake instead of waking it by itself, so a busy channel does not start a run per message.
  • Connections that need attention: a connection it was given is about to expire, has expired, or is due for a new key.

• Talk to it yourself. Pick one of its channels and your own address there (your Slack member ID, your email address). Your messages on that channel join its conversation, and it answers you there. Nobody else’s messages do. On an email channel there is one more choice, off by default: Treat email from my address as me. Anyone can put your address on an email, while Slack and Teams messages can’t be faked that way, so with it off your emails reach the agent like anyone else’s.

• What it may do on its own.

  • Looks things up, and asks you before it changes anything (the default). On its wakes, every tool that may change something waits for your OK in Approvals before it runs.
  • Does things, and asks you first only before what you pick. Switching to this starts the list with every tool that may change something; take off what it may do alone. A tool on this list waits for your OK on every run of the agent, not only its wakes: Try it, a chat on any of its channels (yours or anyone’s), a schedule, and a call through the API or a gateway. The list holds even while Always on is off.

  Your approval rules (for example “ask before a refund over 500”) apply either way.

• Reports. A channel to post reports to, and when it reports: after every wake, only when it did something, or once a day, a short summary of what it did (see The daily summary). Reports also show on the agent’s page and in your notifications. When it is waiting for your OK, it says so there too.

The card then shows what wakes it, what it may do, when it reports, when it last woke and why, and when the timer fires next. Wake now wakes it straight away.

The daily summary

With Once a day, a short summary of what it did, it no longer reports after each wake. Once a day, at the time you pick, it sends one short message about the last 24 hours instead:

• how often it was woken, and by what: its timer, a webhook, a message, you;
• how its work ended: finished, stopped before finishing, or still going;
• what it changed (the tools it used that change something), or that it only looked things up;
• what is waiting for your OK, with a link to Approvals.

A day it did nothing sends nothing.

The summary goes to the channel you picked and to your notifications. To get it by email, turn on email for agent reports in your notification settings. Two things still come straight away: its answers to your own messages, and a note when it is waiting for your OK.

You pick the time and the time zone next to the choice, for each agent. Until you change them, it goes out at 9:00 in your own time zone.

What happens at a wake

1. Everything that woke it since it last worked is collected, oldest first.
2. If it is still working on something, the new things join that work at its next step (“While you were working: …”). It never runs twice at once.
3. Otherwise it starts on its conversation again, with its standing instructions and the list of what happened.
4. It works with its own run limits (steps, cost, time), the same ones you set under Run limits.
5. When it finishes, it answers you where you wrote, and reports where you asked.

When it stops on its own

The agent page says why, and you get a notification:

• It woke more often in an hour than it may. Something kept waking it: a webhook that fires too often, or a channel it reports into and also listens to. Look at What woke it lately, fix the cause, and turn it back on.
• Its owner can no longer run it. Always-on work answers to the agent’s owner, checked at every wake. Add them back, or duplicate the agent.
• It acts as itself, and your plan no longer includes that. An agent set to act as itself (Business) runs with its own access. If the plan lapses it pauses instead of running as you. Switch it back to acting as you, or upgrade, then turn it back on.
• Its timer could not be restored after a restart. Turn it back on.
• Your plan has room for fewer always-on agents on almyty-hosted machines than were on. This only concerns agents that live on an almyty-hosted machine; one on your own machine, or with no machine, is never paused for it. It happens when a plan changes or lapses. The ones turned on last pause first, at their next wake, and the message says how many your plan includes. They turn back on by themselves as soon as there is room: when you turn Always on off for another of them, or your plan includes more. Once there is room you can also turn one back on yourself.

Limits

These are settings, not fixed numbers: an install changes them for any plan (ALWAYS_ON_PLAN_CAPACITY), and an organization can set tighter ones for itself. On machines you run yourself or with no machine at all, the number of always-on agents is not limited on any plan, and changing plans never pauses one. The included count is for almyty-hosted machines when they arrive.