Planning Systems

Introduction

A planning system generates, validates and revises plans. Beyond a single planner LLM, production systems use plan caches, plan templates and constraint checkers.

Beginner analogy: a project manager combines templates, past projects and current constraints — not a fresh blank plan every time.

Understanding the topic

Core concepts:

• Plan templates for repeated goals (cheap, fast).
• Constraint checker rejects illegal plans pre-execution.
• Plan cache: reuse plans for similar goals.
• Plan critic: second LLM scores the plan.
• Promote frequent plans into skills.

Syntax reference

Visual workflow / architecture:

┌──────────┐  goal   ┌──────────┐
│  Cache   │◄────────┤ Router   │
└────┬─────┘         └────┬─────┘
     │ hit               │ miss
     ▼                   ▼
  reuse plan         Planner LLM
                        │
                        ▼
                   Critic / Constraint
                        │ ok
                        ▼
                    Executor

Real-world use

Devin caches plans; LangGraph supports plan-and-execute; Pydantic AI ships planner+critic templates.

Best practices

• Cache plans by goal hash.
• Cheap-first: try template → cache → LLM.

Common mistakes

• Always invoking the planner LLM — slow + expensive.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Three ways to make planning cheaper.
• Q2. What's a plan template?
• Q3. Scenario: 60% of your goals are duplicates. How do you exploit that?