Agent loop

When you will meet it

The AI-agent entry gives you four parts — model, context, tools, loop. This entry is how the fourth part turns; it is the single most important mechanism on the site. Without it you credit the model for every magical behaviour and blame it for every failure, when both usually live in a link of the loop: context assembled wrong, output misparsed, stop condition missing.

Minimal example

while True:                          # 這個迴圈就是 Agent 本身
    ctx = 系統提示 + 對話歷史 + 工具結果   # ① 組上下文
    out = llm(ctx)                     # ② 呼叫模型（唯一的 AI 步驟）
    if out 是純文字:
        回給用戶; break                 # ⑤ 停止條件
    result = run(out.tool_call)        # ③ 解析請求 ④ 執行工具
    對話歷史 += [out, result]           # 結果接回，進入下一輪

Only step ② is the model; ①③④⑤ are ordinary code. Every other agent term you will meet — ReAct, planning, checkpoints — modifies some part of this loop, so learn the loop first.

What people get wrong

• Believing the model itself runs the loop. The model is invoked once per turn and has no idea a loop exists; the surrounding code feeds results back and asks again.
• No stop condition. Without exits for done and give-up, the loop burns tokens forever or oscillates between two actions.