What Is an Agent Turn?

What is an agent turn

An agent turn is everything that happens between a user message and the agent's reply.

That sounds trivial, and for a plain chatbot it is. You send "hello", the model answers, the turn is over. One turn, one model call.

Agents are different. Ask an agent to "find 50 prospects and enrich them with work emails" and it will call a search tool, read the result, call an enrichment tool, read that result, maybe fix a mistake, and only then answer you. All of that is still one turn. The turn ends when the agent stops calling tools and hands control back to you.

A turn is not one model call

Inside a turn, the agent runs a loop. Each pass through the loop is usually called a step:

let messages = [userMessage];

while (true) {
  const response = await llm.complete(messages);

  if (response.toolCalls.length === 0) {
    return response.text; // turn is over
  }

  for (const call of response.toolCalls) {
    const result = await runTool(call);
    messages.push(result); // the model sees this in the next step
  }
}

One turn can contain one step or fifty. The model decides when it is done, which is why agent frameworks put a cap on steps per turn. Without a cap, a confused agent loops forever.

Every step appends messages. The context grows with each tool result, so a long turn is also an expensive turn.

Turn, step, conversation

Different frameworks use different words for the same thing. The Vercel AI SDK calls steps "steps" and marks them with step-start parts in the message stream. Some agent APIs say "iterations". The concept is the same everywhere.

Why the difference matters

💰 Billing. You pay per token, and a turn bills every step inside it. A one-question turn can cost fifty times more than another one-question turn.
⏱️ Limits. Step caps and timeouts apply per turn. If your agent hits them, the work inside the turn is what you need to shrink.
🖥️ UI. While a turn runs, the user sees nothing unless you stream the steps. Showing tool calls as they happen is what makes an agent feel alive.
✋ Interruptions. A user message that arrives mid-turn either waits for the turn to finish or gets injected into the loop. You have to pick one.

Turns in practice

The pipe0 agent is a normal example of this loop. One user message like "find 50 Head of Engineering prospects at robotics companies" becomes a turn with several steps: pick a search, fetch its schema, validate the payload, create a sheet, run the enrichment. The reply comes after the last tool result, and the whole turn is visible as tool calls in the thread.

If you want to watch turns run against real tools, connect an MCP server to Claude Code and give it a task. We wrote up a full session in Claude Code for sales prospecting.