Skip to main content
INFORMATIVEACTIVEDocumentation Governance
Truth Source: Repository schemas and tests are authoritative.

FLOW-04: Single Agent with LLM Enrichment

Source of Truth: tests/golden/flows/flow-04-single-agent-llm-enrichment/

Purpose​

AEL (Action Execution Loop) integration. Validates LLM-enriched plan generation where steps involve LLM execution.

Scope​

This evaluation scenario validates:

  • Plan steps with LLM executor roles via agent_role
  • LLM roles combined with tool and agent roles
  • Protocol-level representation of LLM-powered workflows
  • Plan structure validation (not actual LLM execution)

Non-Goals​

This scenario does NOT evaluate:

  • Basic 2-step flows (see FLOW-01)
  • Large plan volumetrics (see FLOW-02)
  • Tool integration only (see FLOW-03)
  • Multi-round approval (see FLOW-05)
  • Actual LLM API calls (outputs are pre-determined)

L2 Modules Exercised​

ModuleRole in Flow
ContextEstablishes problem domain for LLM-powered workflows
PlanDefines steps with LLM executor roles

Key Protocol Fields​

Plan (LLM Steps)​

  • steps[]: 6 steps with tool/LLM/agent pattern
  • Each step:
    • step_id: UUID v4
    • description: Task description
    • agent_role: Executor type
    • dependencies: Optional sequencing

LLM Roles (via agent_role)​

  • agent_role: "llm_claude" → Claude LLM execution
  • agent_role: "llm_gpt" → GPT LLM execution
  • agent_role: "curl_executor" → HTTP tool (from FLOW-03)
  • agent_role: "jq_processor" → JSON processing
  • agent_role: "agent" → Standard agent logic

Integration Dimensions (L3/L4)​

LLM Integration (via Runtime)​

  • LLMs are runtime capabilities, not L2 protocol constructs
  • agent_role provides protocol-level indication of LLM usage
  • Actual LLM invocation happens at L4 (Runtime layer)
  • For golden tests: LLM outputs are pre-determined

FLOW-04 focuses on protocol-level representation, not actual LLM execution.

Evidence​

TypeLocationStatus
Golden Flowtests/golden/flows/flow-04-single-agent-llm-enrichment/✅ Passed
Input Fixturestests/golden/flows/flow-04-single-agent-llm-enrichment/input/Available
Expected Fixturestests/golden/flows/flow-04-single-agent-llm-enrichment/expected/Available

Expected Behavior​

  • Plan with 6 steps using Tool → LLM → Agent pattern
  • All agent_role values are valid non-empty strings
  • No fictional fields (no llm_model, enrichment_metadata, etc.)
  • Context remains immutable throughout

Invariants Tested​

  • All step_id values unique UUID v4
  • All description fields non-empty
  • All agent_role fields non-empty
  • Step status valid enum values
  • Minimum 6 steps in plan

Document Status: Informative (Evaluation Scenario)
Source of Truth: tests/golden/flows/flow-04-single-agent-llm-enrichment/README.md