ad

agentgraph/methods/production/openai_structured_extractor.py

@@ -88,6 +88,11 @@ WORKFLOW PATTERNS:
 - Simple (1-2 agents): Single consolidated task, basic relations
 - Complex (3+ agents): 3 high-level specialized tasks with NEXT relations (avoid over-decomposition)
 
+ORDER VS HIERARCHY (MANDATORY):
+- Use NEXT (Task→Task) for sequential pipelines where one task hands off to the next. If there are N tasks in a linear flow, expect exactly N-1 NEXT edges to cover the chain.
+- Use SUBTASK_OF (Task→Task) only for hierarchical decomposition where a parent task contains child tasks executed within its scope. Do NOT use SUBTASK_OF for simple multi-stage pipelines without nesting.
+- When both structures exist: use SUBTASK_OF for containment and NEXT among siblings to express order within the parent.
+
 JSON PARSING FOCUS:
 Locate "observations", "agents", "data" sections. Count DISTINCT agents (not mentions).
 Look for *_Expert patterns and execution evidence.
@@ -221,7 +226,7 @@ ANALYSIS STEPS:
      * Output→Human (DELIVERS_TO): 1 relation only
      * Agent→Tool (USES): Include exactly one when a Tool exists
      * Tool→Task (REQUIRED_BY): Include one only when the task explicitly depends on a Tool (execution/platform dependency). Otherwise omit.
-     * Task→Task (SUBTASK_OF): Use sparingly when the trace explicitly states that one task is a component inside another (e.g., Decomposition inside Question Answering). Do not replace NEXT; both may coexist when justified.
+     * Task→Task (SUBTASK_OF): ONLY when the trace explicitly states hierarchical containment (e.g., Decomposition is a component of Question Answering). Never substitute it for linear stage order.
    
    - FORBIDDEN: ASSIGNED_TO, INTERVENES
    - TARGET: Maximum 9–10 total relations for 3-agent workflows