Chain agents

Core has no pipeline object. To run agents one after another, a Flow listens for one run to finish and starts the next. Core owns each run. Flow owns the order.

flowchart LR
    A[Run 1: triage] -->|Agent Lifecycle Event: Succeeded| F{Flow checks the run's status}
    F -->|Succeeded| B[Run 2: draft reply]
    F -->|Anything else| S[Stop]

Build the chain

1. Create a platform event-triggered Flow on Agent Lifecycle Event (AgentLifecycleEvent__e).

2. Add a Decision: Transition__c equals Succeeded and DefinitionId__c equals the first agent’s Agent Definition Id.

3. Add Get Agent Run Result with Run ID set to RunId__c. Add a Decision that continues only when Status is Succeeded. The event is a hint. The run’s own status is the proof.

4. Add Start Agent Run for the next agent:

   Input	Value
   Parent Run ID	{!$Record.RunId__c}
   Correlation Key	{!$Record.RunId__c}-draft
   Anchor Record ID	The same record as the first run
   Input	What the next agent should do. Pass record Ids, not values.

5. Set the user the Flow runs as. Give that user AgentKernelRunAgents, a share of the next agent, and use of its identity’s Named Credential.

Pass Ids, not values, because a run’s sensitive values are tokens that work only inside that run.

Questions

Question	Answer
Whose access does the next run get?	If the first run was limited to a person, the next run keeps that limit, whoever starts it. Otherwise the execution user’s access.
What if a step fails?	The chain stops. Rescue the failed run. When it succeeds, the Flow starts the next step.
What if the event is lost?	Lifecycle events are best effort. The run story shows a parent with no child. Start the next run again with the same Correlation Key.
What does a chain cost?	Each step is a full run: about 2 jobs and 1 API call to start, plus its model and tool steps. Use chains for background work, not for a person waiting on a page.