Planning Pattern

Overview

Planning agents decompose complex tasks into structured roadmaps before execution. They analyze requirements, identify dependencies, and sequence operations logically.

Planning Process

Task Analysis → Subtask Identification → Dependency Mapping → Sequencing → Execution

Plan Structure

{
  "goal": "Create comprehensive market report",
  "subtasks": [
    {
      "id": "1",
      "action": "gather_financial_data",
      "dependencies": [],
      "priority": "high"
    },
    {
      "id": "2",
      "action": "analyze_competitors",
      "dependencies": [],
      "priority": "high"
    },
    {
      "id": "3",
      "action": "synthesize_insights",
      "dependencies": ["1", "2"],
      "priority": "medium"
    },
    {
      "id": "4",
      "action": "write_report",
      "dependencies": ["3"],
      "priority": "low"
    }
  ]
}

Planning Strategies

Hierarchical Task Network (HTN)

Decompose high-level goals into progressively detailed subtasks.

Dependency-Aware Planning

Identify which tasks depend on others and optimize parallel execution.

Adaptive Planning

Revise plans based on execution outcomes (combines with ReAct).

ReWOO: Reasoning Without Observation

ReWOO separates planning from execution entirely:

1. Planning Phase: Generate complete plan without tool access
2. Execution Phase: Execute all tools in planned order
3. Synthesis Phase: Combine results

This reduces LLM calls compared to ReAct but requires confident planning.

When to Use Planning

Good fit:

• Multi-step tasks with clear dependencies
• Research or analysis projects
• Tasks where order matters
• Situations requiring resource optimization

Caution:

• Plans can become outdated quickly
• Over-planning delays execution
• Some tasks are better explored adaptively

Planning + Execution Patterns

Approach	Planning	Execution	Feedback
Plan-then-Execute	Upfront	Sequential	None
ReWOO	Upfront	Parallel	Post-hoc
ReAct	Incremental	Interleaved	Continuous
Adaptive Planning	Initial + Revisions	Monitored	Continuous