Conductor Design Document

Overview

Conductor is the intelligent task scheduler in the AI Kernel Generator. It inherits from AgentBase and manages the entire task execution flow based on workflow.yaml configuration. It uses LLM intelligent analysis of Agent execution results to decide the next execution flow and provide guidance.

Core Functions

• Intelligent Flow Scheduling: Dynamically selects the next Agent based on workflow.yaml configuration and LLM intelligent analysis
• Execution State Management: Records and tracks all Agent execution results, maintaining complete task traces
• Error Handling and Retry: Intelligently handles Agent execution failures with automatic retry mechanisms
• Flow Control: Manages execution steps and repeat limitations to avoid infinite loops

Initialization Parameters

Parameter Name	Type/Required	Description
op_name	str (Required)	Kernel name
task_desc	str (Required)	Task description
task_id	str (Required)	Task ID
dsl	str (Required)	DSL type: "triton_cuda", "triton_ascend", "swft", etc.
framework	str (Required)	Frontend framework: "mindspore", "torch", "numpy", etc.
arch	str (Required)	Hardware architecture: "ascend910b4", "a100", etc.
workflow_config_path	str (Optional)	Workflow configuration file path
config	dict (Required)	Complete configuration dictionary

Workflow Configuration

Conductor manages execution flow based on workflow.yaml configuration files. For detailed configuration documentation, see Workflow Configuration.

Configuration Elements

• agent_info: Agent flow definitions and output formats
• start_agent: Starting Agent
• limitation_info: Execution limits (maximum steps, repeat limits)
• mandatory_llm_analysis: List of Agents requiring mandatory LLM analysis

Configuration Examples

Refer to detailed configuration examples:

• config/default_workflow.yaml - Standard Designer→Coder→Verifier flow
• config/coder_only_workflow.yaml - Simplified Coder+Verifier only flow
• config/conductor_connect_all_workflow.yaml - Fully connected Agent flow

Key Methods

get_next_agent() - Intelligent Decision Process

Executes four decision phases:

1. State Update: Increment step count, clear historical suggestions
2. Retry Check: Handle Agent parsing failure retry logic
3. Option Retrieval: Get valid options based on workflow configuration and current state
4. Intelligent Decision: Call LLM or execute directly based on option count and mandatory analysis configuration

record_agent_execution() - Execution Recording

Records Agent execution results and updates task state:

• Save raw data to execution trace
• Parse results using appropriate parsers
• Update task information dictionary

_llm_decide_next_agent() - LLM Intelligent Analysis

Uses LLM to analyze current execution state and decide the next Agent:

• Build analysis prompts based on conductor/analyze.j2 template
• Comprehensively consider Agent results, error logs, valid options, etc.
• Return decided next Agent name and suggestions

Typical Execution Flow

Based on default configuration:

designer → conductor(decision) → coder → conductor(decision) → verifier → conductor(intelligent analysis) → finish/coder

Custom Extensions

Users can extend Conductor through the following approaches:

1. Configuration Extension: Modify workflow.yaml to define custom Agent flows
2. Code Extension: Override get_next_agent() to add custom decision logic
3. Template Extension: Modify conductor/analyze.j2 to customize LLM analysis prompts