Triggers

A published, active agent can be started in four ways. All of them trigger the same definition — there are no different kinds of agent.

Trigger	Who starts it	When to use
Manual	A person, from Studio or over the API	On-demand runs
Schedule	The clock	Periodic routines
HTTP address	Another system	Integration
Another agent	A flow	Composition

Every trigger points to an exact version

A trigger points to a specific active version, never to "the latest". Publishing a new version does not by itself change what a schedule runs — what happens is that the previous version stops being active.

This is deliberate: a schedule that switched versions on its own would start running, at three in the morning, something nobody tested.

Manual

Trigger it from Studio, filling in the input data on a form built from what the agent declares.

You can also trigger it over the API — useful for scripts. See API.

There is a limit on manual triggers per person per minute. It exists to prevent floods, not to get in the way of normal use.

Schedule

See its own page: Schedules.

HTTP address

See its own page: HTTP addresses.

A call from another agent

Configured inside the calling flow, through the Call another agent step. See Calling another agent.

Differences by environment

	development	production
HTTP address	Always temporary (expires)	Can be permanent
Schedule	Always expires	Durable
Role to create one	People who build	People who publish

A forgotten test in development stops on its own. A production trigger is a decision for whoever answers for it.

Re-running

Re-running creates a new run, from the same active version. It does not reuse the previous one — that history is immutable.

You can re-run with the same input or provide different input. Re-running a production run requires a publishing role.

How the trigger is recorded

Every run records three things separately:

Axis	Example
Who asked	A person, or the run that called it
How it was triggered	Manual, schedule, address, another agent, re-run
Which identity it ran under	The organization's service account

A re-run is recorded as its own type, not as a new trigger: mixing the two would hide a mass re-run in the metrics.

Next steps

• Schedules
• HTTP addresses
• Limits