Agent Loop · Know When Execution Must Stop

01 · 先看一个会卡住的助理01 · Start with a stuck assistant

如果助理不断查到同一个空结果，它会自己停下来，还是一直烧掉预算？If an assistant keeps finding the same empty result, will it stop—or keep burning the budget?

02循环驾驶舱：一次只走一个阶段Loop Cockpit: Advance One Stage at a Time

步进模拟型Step simulation

唯一主动作The one main action

每次点击只执行观察、计划、动作或结果中的一个阶段。默认任务不会完成，预算到顶时停止门自动触发。Each click executes exactly one stage: observe, plan, act, or result. The default task stays unresolved so the budget guard must stop it.

固定预算Fixed budget8 credits

重复阈值Repeat limit2×

初始种子Initial seed20260722

教学模拟值，非真实模型输出simulated for teaching — not real model output

当前圈数Current cycle

—

预算剩余Budget left

credits

连续重复Repeat streak

0×

阈值 2×limit 2×

01 · OBSERVE

观察Observe

—

02 · PLAN

计划Plan

—

03 · ACT

动作Act

—

04 · RESULT

结果与校验Result & check

—

03把一圈拆成 4 个零件Break One Cycle into Four Parts

顺序不可跳过Order matters

PART 01

观察状态Observe state

读取环境此刻真正发生了什么，而不是沿用上一轮猜测。Read what the environment actually says now instead of carrying forward an old guess.

PART 02

选择计划Choose a plan

根据新观察选下一步；计划可以被结果推翻，不是提前写死的承诺。Pick the next step from the new observation. A plan is revisable, not a promise.

PART 03

执行动作Take action

调用工具会消耗预算，也可能带来副作用，因此每次动作都要记账。Tool calls spend budget and may have side effects, so every action must be accounted for.

PART 04

校验结果Check the result

动作发出不等于目标完成。结果决定继续、成功、失败还是强制停止。Sending an action is not task completion. The result decides whether to continue, succeed, fail, or stop.

04最小规则：只有三道门都放行才继续Minimal Rule: Continue Only While All Three Gates Pass

AND gate

go_t+1 = ¬done_t ∧ cost_t < B ∧ repeat_t < R

B = 8 credits，R = 2×。上方任一门变成假，下一圈就不会开始。B = 8 credits and R = 2×. If any gate becomes false, the next cycle never starts.

结果未完成Result unresolved 目标门仍放行Goal gate stays open

检查预算与重复Check budget and repeats 每圈结果后计算Evaluated after each result

继续或停止Continue or stop 停止原因写入日志Reason is written to the log

05边界实验：注入一次重复陷阱Boundary Test: Inject a Repetition Trap

主动制造失败Create a failure

让 Agent 连续重查同一空结果Make the Agent repeat the same empty query

点击后会重置模拟，连续执行同一个工具动作。第二次重复时，重复门会在预算耗尽前自动停止循环。This resets the simulation and repeats one tool call. On the second repeat, the repetition guard stops the loop before the budget runs out.

尚未执行。当前重复阈值固定为 2×。Not run yet. The repeat limit is fixed at 2×.

✗ 工具调用成功，任务就成功了。A successful tool call means the task succeeded. → ✓ 必须用真实结果再次校验目标。The actual result must be checked against the goal.

✗ 只要还有下一步，循环就应继续。If another action exists, the loop should continue. → ✓ 预算与重复门可以否决下一步。Budget and repetition guards can veto the next action.

✗ 多重试几次总会得到不同结果。More retries will eventually produce a different result. → ✓ 无新观察的重复动作只会累积成本与风险。Repeating without new evidence only adds cost and risk.

06 · 一句话带走06 · One line to keep

Agent 循环把观察、计划、动作与校验反复串起来，但只有完成门、预算门和重复门共同放行，下一圈才能开始。An Agent loop repeats observe, plan, act, and check—but another cycle starts only when completion, budget, and repetition gates all allow it.