VISUALBEGINNER
AI Agent Architecture Map
One diagram of how the pieces of a modern agent fit together — loop, tools, memory, evaluation, and the guardrails between them.
The whole system on one page. Everything below the loop is optional at first and mandatory eventually.
┌──────────────┐
user input ────▶│ │
observations ────▶│ AGENT LOOP │◀──── tool results
environment ────▶│ │
└──────┬───────┘
│
┌──────────────┬───────┴───────┬──────────────┐
▼ ▼ ▼ ▼
┌────────┐ ┌──────────┐ ┌───────────┐ ┌────────────┐
│ TOOLS │ │ MEMORY │ │ GUARDRAIL │ │ EVALUATION │
└────────┘ └────┬─────┘ └───────────┘ └────────────┘
│
┌──────────┬───┴────┬────────────┐
▼ ▼ ▼ ▼
short- long-term episodic semantic
term (vector) (log) (facts)Read it as three questions. What can it do (tools). What can it remember (memory). How do you know it worked (evaluation). Guardrails sit between the loop and the world because that is the only place they can actually stop anything.
The part most people skip is the episodic log, and it is the only branch that can answer “why did it do that” after the fact.