Automations
An automation is the same agent you use in a conversation, started another way. A schedule can start it at a set time. An event in a connected app can start it when something happens.
The agent brings the same job, instructions, skills, tools, permissions, and files. The difference is that no one opens the chat and types the first message.
A schedule can start a familiar job
Suppose the marketing launch agent prepares a campaign report with you every Monday. It reads the latest analytics, explains what changed, and posts a short summary to the team channel.
Once the report follows a stable pattern, you can schedule it for Monday morning. The schedule supplies the message the agent receives on each run, for example:
Summarize last week's Team Workspaces campaign. Compare results with the previous week, explain the two largest changes, and post the report to the launch channel.
The agent handles the report with the same working setup it uses in the playground. The schedule only supplies when to begin and what to do.
An app event can start work when something changes
Some jobs should begin when an event happens rather than at a particular time.
The launch agent could start when a campaign changes to completed, when a new row appears in the launch tracker, or when a review issue receives a particular label. The event carries details about what happened, and the trigger can include those details in the message the agent receives.
This makes the work responsive. The agent does not need to check every hour for a change; the connected app starts the run when the relevant event arrives.
An unattended run needs a complete starting point
In a chat, the agent can ask what you meant by "the campaign," which date range to use, or where to put the result. An automation should begin with those choices already made.
Prepare three layers of context:
- The agent's instructions hold facts and rules that apply to all of its work.
- The trigger message describes this recurring job and includes the details supplied by the schedule or event.
- The agent's files and skills provide the reference material and task process it may need.
For the Monday report, the instructions describe the company and audience, the reporting skill describes the analysis process, and the scheduled message names the campaign, comparison period, and destination.
The agent needs permission to finish the job
An automated run has no active playground conversation where someone can answer an approval card. If a required tool waits for approval, the report cannot reach its destination on its own.
Before activating the automation, review the exact actions it needs. The launch agent may need permission to read analytics and post to one team channel, but it does not need every available action in those apps.
Set required actions to run without asking, deny actions the automation should never take, and remove tools that do not belong to the job. See Permissions for the model and Control what an agent can do for the settings.
The result needs a destination
When you work in the playground, the answer appears in front of you. An automation should leave its result somewhere your team will see it.
That destination might be a channel, an email, a document, a task, or a row in a sheet. Give the agent the appropriate tool, name the destination in the trigger message, and decide what should happen if delivery fails.
If later runs should use the result, the agent can also save a copy in its shared files. Monday's report can then compare itself with the report from the week before.
Each automation runs a version you choose
A schedule or event trigger points to a committed agent version. Draft changes in the playground do not silently change an automation that is already running.
This gives you a clear review point. Improve the agent, test the recurring message in the playground, commit the new version, and then choose when the automation should use it.