Skip to content

Workflows & Workflow Sessions

This feature covers declarative multi-agent graph workflows, DAG state machine transitions, and execution sessions.

1. Declarative Multi-Agent Workflows (AiWorkflow)

An AiWorkflow defines a Directed Acyclic Graph (DAG) or state machine template composed of nodes and conditional edges.

1.1 Node Types

graph TD
    AGENT[AGENT Node<br/>Executes AiAgent with prompt template]
    CONDITION[CONDITION Node<br/>Evaluates SpEL expression on context]
    APPROVAL[HUMAN_APPROVAL Node<br/>Pauses execution for human signal]
    TOOL[TOOL Node<br/>Invokes direct tool/MCP call]
  • AGENT: Executes an AiAgent with a prompt template, saving output into workflow memory under outputKey.
  • CONDITION: Evaluates a Spring Expression Language (SpEL) expression against workflow context (e.g., #data['research_data'] != null).
  • HUMAN_APPROVAL: Transitions session state to WAITING_APPROVAL and pauses Temporal execution until a human approval signal is received.
  • TOOL: Directly executes a skill or MCP tool step.

2. Workflow Sessions (WorkflowSession & WorkflowExecutionService)

A WorkflowSession represents an instance of a workflow execution: - GitOps Trigger: Created declaratively via WorkflowSession CRD (kubectl apply -f session.yaml). - REST API Trigger: Created imperatively via POST /api/v1/workflows/{name}/sessions.

2.1 Session Execution Lifecycle

PENDING ──► RUNNING ──► WAITING_APPROVAL ──► COMPLETED
               │                                │
               └───────────► FAILED ◄───────────┘
  • Loop Guards: Prevents infinite loops via maxLoops parameter (default 10).
  • Session Logs: Audit entries logged per step in workflow_session_logs table.