Workflow System Design Document

Overview

Workflow is AIKG's core configuration system that defines Agent execution flow through YAML configuration files, provides intelligent scheduling rules for Conductor, and supports flexible workflow customization.

Configuration File Structure

Basic Format

# Agent information definition
agent_info:
  agent_name:
    possible_next_agent: [next_agent1, next_agent2]
    output_format:
      parser_name: parser_name
      parser_definition:
        output_fields:
          field_name:
            field_type: str
            mandatory: true
            field_description: "Field description"

# Starting Agent
start_agent: agent_name

# List of agents requiring mandatory LLM analysis (when these agents enter conductor, conductor forces LLM analysis)
mandatory_llm_analysis: [agent1, agent2]

# Execution limitations
limitation_info:
  required: # Required limitation information
    max_step: 20  # Maximum execution steps
  optional: # Optional limitation information
    repeat_limits:
      single_agent:
        agent_name: 3                 # Maximum consecutive repeats for single Agent
      sequences:
        sequence_name:
          pattern: [agent1, agent2]   # Sequence pattern
          max_repeats: 3              # Maximum sequence repeats

Core Configuration Fields

Field Name	Type	Required	Description
agent_info	dict	Yes	Define configuration information for all Agents
start_agent	str	Yes	Specify starting Agent
mandatory_llm_analysis	list	No	List of Agents requiring mandatory LLM analysis
limitation_info	dict	Yes	Execution limitation configuration

Predefined Workflow Types

Configuration files reference: python/ai_kernel_generator/config/ directory.

1. Default Workflow (default_workflow.yaml)

Flow: designer → coder ←→ verifier → finish

• Complete design→coding→verification flow
• Can rollback to coder when verification fails

2. Fully Connected Workflow (conductor_connect_all_workflow.yaml)

Flow: Supports flexible transitions between all Agents

• Maximum flexibility, Agents can self-repair

3. Coder-Only Workflow (coder_only_workflow.yaml)

Flow: coder ←→ verifier → finish

• Skip design phase, directly generate code

4. Verifier-Only Workflow (verifier_only_workflow.yaml)

Flow: verifier → finish

• Minimalist flow for verification only

Agent Configuration Description

Agent Information Structure

agent_name:
  possible_next_agent: [list]         # List of possible next Agents
  output_format:                      # Output format definition (optional)
    parser_name: str                  # Parser name
    parser_definition: {...}          # Parser definition

Decision Mechanisms

For detailed decision logic implementation, see Conductor Design Document.

Automatic Decision

• No available Agents → Directly finish
• Single available Agent and not mandatory analysis → Direct execution
• Parsing failed and retryable → Retry current Agent

LLM Intelligent Decision

• Multiple available Agents → Intelligently select optimal path
• Single available Agent but in mandatory_llm_analysis list → Analyze and provide suggestions

Configuration files reference: python/ai_kernel_generator/config/ directory for various workflow.yaml files.