Use Adapt
Automations
An automation runs an agent without anyone typing a message: on a schedule, on demand, or when an event arrives on one of the agent's surfaces.
Automations belong to an agent and live on its Automations tab. Each one has instructions (what the agent should do) and a trigger (when it runs). Any member can add scheduled automations to any agent they can see. Event automations need permission to edit the agent.
Create an automation
Open Agents, pick the agent, and stay on its Automations tab. Then either:
- Describe it in a sentence, such as "Every Monday at 9:00, summarize last week's support tickets". The agent's model reads it back as a filled-in automation: a name, a schedule, instructions, and where results go. Check it and change anything that's off.
- Choose Create from scratch and fill in the fields yourself.
Agents can also suggest automations in a web chat when the conversation shows work you'll likely want repeated. Nothing is set up until you choose Set up on a suggestion.
Every automation has three parts:
- When: a schedule, Just once, or When something happens.
- Does: the instructions the agent gets on each run, up to 32,000 characters.
- Then: where the result goes.
Choose Turn on to save it. If the agent needs an app it isn't connected to, Adapt lists it with a way to connect it, or you can choose Turn on anyway.
Run on a schedule
Pick how often it repeats: every 15 minutes, every 30 minutes, every hour, every day, every weekday, every week, every month, or every year. Daily and longer schedules take a time and a start date; every schedule takes a timezone. Weekly, monthly, and yearly repeats use the start date's day, and months without that date are skipped. Just once takes a date and time in your own timezone.
- Repeats are at least 15 minutes apart.
- A one-time run must be in the future when you save it. It runs once, then shows as Completed and can't be edited.
- If Adapt was down when runs were due, it makes one overdue run and then continues from the next scheduled time. Missed runs aren't replayed one by one.
Each scheduled run starts its own conversation, so runs that overlap don't share one. The agent needs a model for scheduled runs to start.
Choose where results go
Every scheduled run keeps its result in its own conversation, which you open from the automation's run history. Automation conversations don't appear in the chat list.
To also post the result elsewhere, attach a surface to the agent first; Then then offers it as a destination:
- Slack: a channel, chosen from channels you can read.
- Email: an email address.
- iMessage and SMS: a phone number.
Retries and end dates
Open Details in the editor:
- Retries: Don't retry, or retry a failed scheduled run up to 5 times. Retries wait 1, 2, 4, and so on minutes after the failed run ended.
- Ends (repeating schedules only): Never, After a number of runs (1–10,000 scheduled runs; retries and Run now don't count), or On a date (the last day it runs, in your timezone).
Run, pause, and edit
Select an automation to open it. It shows its status, who it runs as, its schedule, its instructions, where results go, its next run, and its recent runs.
- Run now starts a run immediately, outside the schedule. It's available for scheduled automations only.
- Pause stops future runs until you choose Resume.
- Edit changes the automation.
- Delete removes it. Its past results stay in their conversations.
Pausing or saving a change cancels runs that are queued but haven't started; they show as interrupted. A run already in progress finishes with the instructions it started with.
Each recent run shows as Queued, Running, Succeeded, Failed, or Interrupted, and opens to its conversation. A failed run offers Run again.
Respond to surface events
Set When to When something happens to run the agent on an event from one of its surfaces, then pick the Connected surface and its Events. The agent's answer goes back to the thread the event came from, such as the Slack thread or the GitHub pull request.
Attaching a surface to an agent adds its default automations, which you can edit, pause, or delete:
- Slack: Respond to Slack messages, for mentions, direct messages, and channel messages.
- GitHub: Respond to trusted member mentions, and Review pull requests, which starts paused.
- Email: Respond to email.
- iMessage and SMS: Respond to text messages.
Without a matching automation that's turned on, the agent doesn't answer on that surface. Deleted defaults stay deleted when you save the agent again.
For each selected event:
- Allow responses to non-workspace members is off by default. Off, only verified workspace members get answers, and each run acts as that member. On, other people get answers too; their runs act as no one, without any member's personal connections or permissions.
- Allow responses to bots (Slack and GitHub) is off by default. On, other bots get answers and their runs act as no one. The agent's own messages are always ignored.
- On GitHub, choose which pull request actions to respond to, and which repositories. With none selected, it covers all repositories, including ones added later.
The editor lists the events each surface supports. That list says what Adapt can handle, not what the provider is set up to send: the surface still has to deliver those events. Run now isn't available for event automations, since a run needs a real event. The surface pages cover each one's events: Slack, GitHub, Email, iMessage and SMS. To start a run from your own program, see Webhooks.
If a surface's connection is paused, its automations keep their settings but no events run. Removing a surface from the agent deletes the automations that use it.
Who can manage an automation
Only the person who created an automation can edit, pause, delete, or run it, whatever their role. Scheduled runs act as that person, with their permissions.
The agent's creator and workspace owners and admins can see every automation on the agent, its run history, and the conversations of its scheduled runs. Other members see only their own automations.
When a member is removed from the workspace, their automations pass to the admin who removed them and are paused. Deleting an agent deletes its automations.
Limits
- Name: up to 200 characters.
- Instructions: up to 32,000 characters.
- Repeating schedules: at least 15 minutes apart.
- Retries: 0–5 per scheduled run.
- Ends after: 1–10,000 scheduled runs.
- Result destinations: up to 10 per automation.