An agent is a while loop with three rules

The word "agent" sounds magical. The implementation is anything but. Claude Code, Cursor, every coding agent you've used — they all run the same five-step loop. Once you see it, you'll never be intimidated by the term again.

1. Send messages + tool definitions to the model.
2. Read the response.
3. If stop_reason == "end_turn": print the text and STOP.
4. If stop_reason == "tool_use": run the requested tool, append the
   tool_result to messages, GOTO 1.
5. (Bound the loop with a max iteration count to prevent runaways.)

That's the agent loop. The model decides what to do; your code wires it to the world.

The three blocks you'll see in content

After the model runs, response.content is a list of blocks. The two that matter for agents:

• {"type": "text", "text": "..."} — natural language for the user
• {"type": "tool_use", "id": "...", "name": "...", "input": {...}} — the model is asking you to run a tool

When you respond back to the model, you send a user turn with a matching tool_result block:

{
    "role": "user",
    "content": [
        {
            "type": "tool_result",
            "tool_use_id": "toolu_01",      # MUST match the tool_use id
            "content": "72°F, sunny",
        }
    ],
}

The tool_use_id link is mandatory. The model needs to know which tool call this result belongs to, because it may have made several in parallel.

The four stop reasons you actually care about

stop_reason	What it means
end_turn	Done. The model is finished. Exit the loop.
tool_use	Run the requested tools and call again.
max_tokens	The response was cut off. Increase budget or summarize.
stop_sequence	A custom stop string was hit (rarely used).

If your loop doesn't check stop_reason, you have two bugs waiting: ignoring tool_use (the agent stalls) and never exiting on end_turn (infinite loop, dead API key).

Where AI specifically gets this wrong

• Treating any response as final. Cursor's first agent attempt almost always misses the tool_use branch and returns whatever text was alongside the tool call as the "answer." Wrong answer.
• Mismatched tool_use_id. Hardcoding "id": "1" in the tool_result instead of echoing the model's actual id. The model rejects the next call.
• No iteration cap. A misbehaving agent that loops on the same tool call can drain a budget in minutes. Always cap at 8-16 iterations.

Browser note: real agent loops need network calls. We'll mimic the model with a hardcoded function that returns either tool_use or end_turn based on the conversation so far. Same loop, no wire.

What everyone calls it

Anthropic's Building Effective Agents names this "the augmented LLM" — a loop that runs until the model says it's done. Every framework wraps the same primitive:

Framework	Primitive
Anthropic SDK	messages.create(...) returns a response with stop_reason ∈ {end_turn, tool_use, ...}. The loop is hand-rolled around it.
OpenAI Responses API	response.output[] blocks (text, tool_call); response.status carries the equivalent of stop_reason.
LangGraph	StateGraph with a ToolNode and a conditional edge that re-enters the model node when tool_calls are present.
Vercel AI SDK	streamText({ tools, stopWhen }) — the loop is implicit; stopWhen controls termination.
OpenAI Agents SDK	Agent + Runner — the runner wraps the loop you'd write yourself.

Same primitive everywhere. Different ergonomics, same stop_reason branching underneath. Once you've written it once (step 8), you'll read every framework's docs and recognize the shape.

Run the editor. We inspect a single tool_use response — what your loop sees right before it has to run the requested tool.