Planning

Planning (sometimes called “Reasoning & Acting”) decouples the strategy from the execution. Instead of reacting immediately to a user request, the agent pauses to decompose the goal into sub-goals, identifies dependencies, and creates an ordered list of steps. This allows agents to tackle complex, multi-step problems that require foresight.

When to Use

• Multi-Step Tasks: “Research X, then Y, then write a report comparing them.”
• Dependency Management: When step B cannot start until step A is finished (e.g., compile code -> run tests).
• Resource Constraints: To optimize the usage of expensive tools or API calls.
• Error Recovery: If a step fails, the plan can be adjusted dynamically without restarting from scratch.

Use Cases

• Travel Itinerary: Searching for flights, hotels, and activities, then checking availability, then booking.
• Software Development: Designing a system -> Writing code -> Writing documentation.
• Data Analysis: Plan -> Data Collection -> Cleaning -> Analysis -> Visualization.

Implementation Pattern

def planning_workflow(goal):
    # Step 1: Create Plan
    # The planner generates a list of steps, not the actual work.
    plan = planner_agent.run(
        prompt="Create a step-by-step plan to achieve this goal...",
        input=goal
    )
    
    results = {}
    
    # Step 2: Execute Plan
    for step in plan.steps:
        # Check dependencies
        if not check_dependencies(step, results):
            raise DepedencyError(f"Cannot execute {step.id}")
            
        # Execute the specific step using a worker agent
        result = worker_agent.run(
            prompt=f"Execute this step: {step.description}",
            context=results # Pass context from previous steps
        )
        
        results[step.id] = result
        
    # Step 3: Summarize
    return synthesizer_agent.run(results)