Agent Autonomy Classification

The Spellcasters: From Novice to Archmage

This is an attempt to provide a classification system for AI Agents. I dislike when simple Chatbots are called agents, so here we are.

Autonomy Levels

Level 1: Scribes

"The Scroll Readers"

  • Powers:

    • Wields 3-7 tools based on specific commands
    • Follows explicit multi-step instructions
    • Maintains context within a single quest

  • By the Numbers:

    • Tools: 3-7 distinct capabilities
    • Independence: 3-5 reasoning steps without guidance
    • Memory: Retains context across single session

  • Examples: Code analyzers, specialized research assistants

Level 2: Conjurers

"The Ritual Casters"

  • Powers:

    • Orchestrates 8-15 tools in sophisticated sequences
    • Develops and follows branching action plans
    • Remembers past interactions to inform current work

  • By the Numbers:

    • Tools: 8-15 orchestrated capabilities
    • Planning: Creates plans with 5-10 dependent steps
    • Autonomy: Completes complex tasks with minimal intervention

  • Examples: Autonomous development assistants, specialized problem-solvers

Level 3: Archmages

"The Reality Shapers"

  • Powers:

    • Summons and coordinates multiple specialized sub-agents
    • Transforms vague intentions into precise execution plans
    • Learns and evolves from each interaction

  • By the Numbers:

    • Orchestration: Manages 3+ specialized sub-agents
    • Depth: 15+ autonomous reasoning steps
    • Capability: Handles tasks requiring 10+ different skills

  • Examples: Full-lifecycle project systems, autonomous research environments

NOT True Agents: The Imposters

Forbidden from the Classification:

  • Pre-programmed Mimics: Customer support chatbots, FAQ responders
  • Connected Scrolls: Systems that merely chain together multiple LLMs without true orchestration
  • Potion Mixers: Low-code/no-code platforms and "building tools" that require human configuration
  • Enchanted Parchments: Workflow tools like Zapier that execute predefined sequences
  • Apprentice Assemblies: Systems requiring extensive setup rather than operating from natural language
  • Separated Spellbooks: Any system requiring multiple separate prompts rather than accepting a single instruction

The True Test: A genuine agent must transform a single natural language instruction into autonomous orchestration of diverse tools without requiring configuration, setup steps, or multiple interaction points. The user casts one spell, and the agent completes the entire quest.