Long Running Agent

Build resilient autonomous agents that can parse PRDs/specifications, generate structured task lists, and execute tasks autonomously over extended periods with state persistence and automatic recovery.

This skill provides agent-agnostic patterns that work with any AI agent framework including Cursor, OpenCode, Claude, and others.

Core Architecture

A long-running agent consists of seven core systems that work with any agent framework:

1. PRD/Spec Processing - Parse requirements documents into structured, executable task lists
2. Task Execution Engine - Autonomous task processing with dependency management
3. API Rotation & Management - Intelligent API key rotation, rate limiting, and load balancing
4. State Management - File-based persistence for workflow states and task tracking
5. Error Handling - Classification, recovery strategies, and graceful degradation
6. Cross-Session Persistence - Resume work across interruptions and restarts
7. Learning & Memory - Pattern recognition and improvement over time

These patterns are framework-agnostic and can be implemented with any AI agent that has file system access.

Implementation Workflow

Step 1: Set Up Project Structure

Create the basic directory structure for persistent state management:

def setup_project_structure(project_name: str):
    """Create directory structure for long-running agent."""
    directories = [
        f"tasks/{project_name}",
        f"results/{project_name}", 
        f"memories/{project_name}",
        f"logs/{project_name}"
    ]
    
    for directory in directories:
        os.makedirs(directory, exist_ok=True)
        print(f"✅ Created: {directory}")

Step 2: Implement PRD Processing

Parse requirements documents into structured, executable task lists:

def parse_prd_to_tasks(prd_content: str, project_name: str) -> Dict:
    """Parse PRD into structured task list with dependencies."""
    # See references/prd-processing.md for full implementation
    
    tasks = {
        "project_name": project_name,
        "created_at": datetime.now().isoformat(),
        "total_tasks": 0,
        "completed_tasks": 0,
        "tasks": []
    }
    
    # Extract sections, analyze dependencies, categorize tasks
    # Returns structured JSON with full task metadata
    return tasks

Full Implementation: See references/prd-processing.md

Step 3: Set Up API Rotation and Management

Configure intelligent API rotation for external service calls:

def setup_api_rotation(api_configs: List[Dict]):
    """Setup API rotation with multiple endpoints."""
    # See references/api-rotation.md for full implementation
    
    global api_manager
    api_manager = APIRotationManager()
    
    for config in api_configs:
        api_manager.add_endpoint(
            name=config["name"],
            base_url=config["base_url"], 
            api_key=config["api_key"],
            rate_limit=config.get("rate_limit", 60),
            quota_limit=config.get("quota_limit", 1000)
        )
    
    print(f"🔄 API rotation configured with {len(api_configs)} endpoints")

Full Implementation: See references/api-rotation.md

Step 4: Implement State Management

Create persistent state management for cross-session continuity:

def save_task_list(task_list: Dict, file_path: str = None):
    """Save task list to persistent storage."""
    # See references/state-management.md for full implementation
    
    if not file_path:
        file_path = f"tasks/{task_list['project_name']}/current_tasks.json"
    
    os.makedirs(os.path.dirname(file_path), exist_ok=True)
    with open(file_path, 'w') as f:
        json.dump(task_list, f, indent=2)

def load_task_list(project_name: str = None, file_path: str = None) -> Dict:
    """Load task list from persistent storage."""
    # Implementation details in references/state-management.md
    pass

Full Implementation: See references/state-management.md

Step 5: Implement Task Execution Engine

Execute tasks autonomously with dependency management:

def execute_next_task(project_name: str) -> Dict:
    """Execute the next available task with dependency checking."""
    # See references/task-execution.md for full implementation
    
    task_list = load_task_list(project_name)
    
    # Find next executable task (dependencies met, status pending)
    next_task = find_next_executable_task(task_list)
    
    if not next_task:
        return {"status": "no_tasks_available"}
    
    # Execute task by category with API rotation support
    result = execute_task_by_category(next_task)
    
    # Update task status and save state
    update_task_status(next_task["id"], "completed" if result["success"] else "failed")
    
    return result

Full Implementation: See references/task-execution.md

Step 6: Set Up Learning and Memory System

Implement pattern recognition and continuous improvement:

def save_execution_pattern(task: Dict, execution_result: Dict, pattern_file: str = "memories/patterns.json"):
    """Save successful execution patterns for learning."""
    # See references/learning-system.md for full implementation
    
    pattern = {
        "task_category": task["category"],
        "task_type": task.get("type", "general"),
        "execution_approach": execution_result.get("approach"),
        "success_factors": execution_result.get("success_factors", []),
        "timestamp": datetime.now().isoformat()
    }
    
    # Save pattern for future reference
    patterns = load_json_file(pattern_file, [])
    patterns.append(pattern)
    save_json_file(pattern_file, patterns)

Full Implementation: See references/learning-system.md

Step 7: Agent Integration

Integrate with your specific AI agent framework:

# For any agent framework (Cursor, OpenCode, Claude, etc.)
def run_long_running_agent(prd_content: str, project_name: str):
    """Main entry point for long-running agent workflow."""
    
    # 1. Setup
    setup_project_structure(project_name)
    setup_api_rotation(load_api_config())
    
    # 2. Parse PRD
    task_list = parse_prd_to_tasks(prd_content, project_name)
    save_task_list(task_list)
    
    # 3. Execute tasks
    while has_pending_tasks(project_name):
        result = execute_next_task(project_name)
        
        if result["status"] == "no_tasks_available":
            break
            
        # Learn from execution
        if result.get("success"):
            save_execution_pattern(result["task"], result)
    
    # 4. Generate summary
    return generate_project_summary(project_name)

Agent Framework Integration

For Cursor, OpenCode, and other AI Agents:

1. Load this skill when starting a new project or resuming work
2. Call run_long_running_agent() with your PRD content
3. Monitor progress through the generated task files
4. Resume anytime by calling execute_next_task()

Example Workflow:

# Start new project
prd = "Your PRD content here..."
summary = run_long_running_agent(prd, "ecommerce-platform")

# Resume existing project  
result = execute_next_task("ecommerce-platform")

# Check status
status = get_project_status("ecommerce-platform")

Key Patterns Summary

File Structure

project-name/
├── tasks/project-name/
│   ├── current_tasks.json    # Current task list and status
│   └── task_history.json     # Completed task history
├── results/project-name/
│   ├── task_001/            # Individual task outputs
│   └── task_002/
├── memories/project-name/
│   ├── patterns.json        # Learned execution patterns
│   └── solutions.json       # Error solutions
└── logs/project-name/
    └── execution.log        # Detailed execution logs

Reference Files

For detailed implementations, see:

• prd-processing.md - PRD parsing patterns, task extraction, structured generation
• api-rotation.md - API rotation, rate limiting, load balancing, error handling
• task-execution.md - Autonomous task processing, dependency management, status tracking
• state-management.md - File-based persistence, cross-session continuity, data integrity
• agent-integration.md - Integration patterns for different AI agents (Cursor, OpenCode, etc.)
• parallel-execution.md - Concurrent task processing, thread safety, coordination patterns
• error-handling.md - Error classification, recovery strategies, graceful degradation
• learning-system.md - Pattern recognition, continuous improvement, memory management

Quick Start

1. Parse your PRD: tasks = parse_prd_to_tasks(prd_content, "my-project")
2. Start execution: run_long_running_agent(prd_content, "my-project")
3. Monitor progress: Check files in tasks/my-project/
4. Resume anytime: execute_next_task("my-project")

Agent Instructions

This skill works with any AI agent that can:

• Read and write files
• Execute Python functions
• Maintain state across conversations
• Handle JSON data structures

Simply load this skill and call the main functions with your PRD content to begin autonomous task execution with full persistence and recovery capabilities.