Anatomy of an agent

Strip away the frameworks and an agent is three things in a loop:

1. A model that decides what to do next.
2. Tools it can call to affect the world or gather information.
3. A loop that feeds tool results back to the model until the task is done.

user goal
   │
   ▼
┌──────────────┐   calls tool   ┌──────────┐
│    model     │ ─────────────▶ │  tool    │
│ (decides)    │ ◀───────────── │ (acts)   │
└──────────────┘   observation  └──────────┘
   │  done?
   ▼
final answer

Everything else is an elaboration of this loop:

• Workflows & orchestration make the loop deterministic where it needs to be.
• Multi-agent systems run several of these loops and let them talk.
• Memory lets the loop remember across runs.
• RAG enriches the model’s context before it decides.
• Evals check that the loop produces the right outcome, repeatably.

See it in practice in the ADK Tutorial, or browse everything in the gallery.