agent-anatomy

Interactive visualization of how LLM agents actually work — the real API message flow between User, Agent, and LLM.

[](https://jegwan.github.io/agent-anatomy/)

<p align="center"> </p>

> Agents aren't magic. They just assemble context. > > system_prompt + tools[] + messages[] → LLM API → if tool_use: execute → append tool_result → repeat. > > That loop is the entire "intelligence" of an LLM agent.

An interactive sequence diagram showing 6 turns of a real agent session with actual JSON payloads:

| Turn | What happens | Key insight | |------|-------------|-------------| | 1 | User request → Agent assembles context → LLM API call | system + tools[] + messages[] — that's all the LLM receives | | 2 | Tool result fed back → LLM decides next action | tool_result goes in as "role": "user" — there is no "tool" role | | 3 | Test failure → self-correction | Error logs in context → LLM can reason about failures | | 4 | MCP tool call | MCP tools are just mixed into tools[] — the LLM doesn't know MCP exists | | 5 | Skill invocation | A skill is just a prompt template injected into the user message | | 6 | Loop termination | No tool_use in response = agent stops the loop |

┌──────────┐     ┌──────────┐     ┌──────────┐
│   User   │ ──→ │  Agent   │ ──→ │   LLM    │
│  (human) │ ←── │ (program)│ ←── │(only AI) │
└──────────┘     └──────────┘     └──────────┘

• User — a human giving instructions
• Agent — a program (deterministic code) that assembles context and executes tools
• LLM — the only AI in the system; it reads context and outputs the next action

• **Step-by-s