Reflex to Learning: The Classic Taxonomy of AI Agent Architectures

Five rungs, one force - each level fixes the one below.

• Left column - The ladder, not a menu:
• Not a menu of seven equal types
• A ladder of increasing demand
• Each rung fixes the level below

• Center column - The five rungs:
• 1. Reflex: rules on now, no memory
• 2. Model: inner state, sees through gaps
• 3. Goal: search when goal is not sensed
• 4. Utility: score trades, expected value
• 5. Learning: improve from experience

• Lower band - The second axis:
• Hierarchical = stack of agents
• Multi-agent = shared environment
• A topology, not a rung - two axes

• Bottom band - Falsify it:
• Test 1: full-view world - model vs reflex
• No measurable edge = decorative line
• Test 2: speed vs safety, no utility - thrash
• Utility resolves it - why rung 4 exists

Sticky note - the one line to keep: Seven types is a menu. Each rung is a fix for the failure below it.