Start a run

From a Flow

Use the Start Agent Run action. It works in record-triggered, scheduled, screen, and platform event-triggered Flows.

Input	Required	What to enter
Agent Developer Name	Yes	The agent’s Developer Name.
Input	No	What the agent should do, as text or JSON.
Anchor Record ID	No	The record the run is about. The agent sees its name and object. Both the starter and the agent must be able to read it.
Correlation Key	No	Recommended in record-triggered Flows. A second start with the same key returns the first run instead of starting another. Example: {!$Record.Id}-triage.
Parent Run ID	No	The run this one continues. See Chain agents.

The action returns the Run ID at once. The run continues in the background. An error, such as an agent not shared with the user, comes back in Error Message instead of failing the Flow.

Who may start a run

A user needs the AgentKernelRunAgents permission set and a share of the Agent Definition. See Give people access.

Whose access the run gets

How the run starts decides this. The starter cannot choose.

Started by	Agent with an identity	Agent without an identity
A person’s click, screen Flow, or record save (including a record-triggered Flow their save fires)	The execution user, limited to what that person can see	The person
A scheduled Flow, scheduled path, platform event Flow, Apex trigger, or background job	The execution user	Refused. Nobody owns this access.

For work that needs the execution user’s full access after a save, start the run from a scheduled path or a platform event-triggered Flow. Give that Flow’s running user AgentKernelRunAgents, a share of the agent, and use of the identity’s Named Credential.

Read the result

Use the Get Agent Run Result action with the Run ID. It returns:

Output	Meaning
Status	Running, Waiting, Paused, Succeeded, Failed, or Cancelled.
Output	The final answer. Sensitive values stay as tokens unless the user has AgentKernelTrustResolver.
Error Code and Error Message	Why a run paused or failed.

A run takes seconds to minutes. To act when it finishes, subscribe a platform event-triggered Flow to Agent Lifecycle Event, as in Chain agents. Always read the run’s status with Get Agent Run Result. The event is a hint, not proof.

From Apex

Call AgentRunService.start. Developers can read a run’s full history with AgentTraceService. See Extend Pluto.

Cost

Start	Cost
Agent with an identity, started by someone other than its execution user	2 jobs and 1 API call
Agent with an identity, started by its execution user	1 job
Each model or tool step	1 job
A typical run with 3 tool calls	About 8 jobs

See Limits and cost.