前段时间刷了 Claude Code 的源码，感觉 Agent Loop 就是这么一套简单的 while 循环：模型返回 Tool 请求 → 执行工具 → 将结果放回消息 → 请求模型；直到模型不再调用 Tool 时结束。代码看起来很朴素，却在生产环境里被塞进了权限确认、上下文压缩、输出截断恢复、模型 fallback、Hook、取消、远程会话、JSONL 持久化等多种能力。

只是当任务不再是一次短暂的终端交互，而是可能跨越进程生命周期、等待外部审批甚至被迁移到其他 Worker 时这种基于内存 Promise 的 while 循环就显得捉襟见肘了。Goose 将 Agent Loop 改造成一个状态机，就是为了解决这类痛点。怎么说呢，

说到使用者痛点。等待审批与长任务的不确定性

• 审批可能来自外部程序： 工具调用往往需要业务审批，而审批流程可能在公司内部程序或第三方网站完成。
• 工具执行可能延迟数小时甚至数天： 数据查询或删除操作需要耗费时间，而且有可能在后台异步完成。
• 进程崩溃或机器休眠导致连接失效： 终端关闭后原来的 Promise 与其 resolve 都会消失，无法继续等待。
• 任务需要横向 或迁移到不同节点： 为了弹性伸缩，执行同一任务的进程可能会变化。
• 缺乏可复现的状态记录： 只依赖内存中的临时变量无法在重启后恢复，需要持久化状态来保证一致性。

Coleade Code 的 While 循环到底能做什么？

Coleade Code 的主循环是一个围绕“本次活跃查询”不断推进的调度中心。它维护了完整的 {messages,toolUseContext,turnCount…} 状态对象。在模型因上下文过长失败时它会尝试压缩消息并更新状态；在输出被截断时也会生成新的状态并继续下一轮请求。其实，整个循环既能处理即时交互，又能在一定范围内恢复错误。但它仍然依赖于持续存在的进程和调用栈。

Coleade Code 示例：即时交互场景

while {
const assistantMessages = await callModel;
const toolUses = findToolUses;if return，const toolResults = await runTools;messages =，}

This loop works well when everything happens within one terminal session and user approvals arrive instantly.

The Goose State Machine：跨进程持久化方法

Goose 把 Agent Loop 展开成一个可重复进入的状态机。每个 “turn” 在到达明确停止点后返回，整个 Conversation 本身即为状态。官方提出将每个 turn 当作普通函数调用，使其易于测试和推理。同时支持事件驱动架构和可组合 Operation。

The Goose 状态机主要流程

loop {
let session = load_session.await?，let Some = machine.step.await?else { break;}，apply_effects.await?，if result.yield_to_client { break;}
}

The state machine itself doesn’t hold long-lived business logic in memory;话说回来，it merely loads current state from persistent storage,decides which Operation applies next based on that state,produces Effects。and n persists those effects.

A Demo Without In-Memory Flags

function nextStep {
// Derive current status purely from history
// ...
}
async function executeTool { /* async call */ }
async function advance {
const event = nextStep;if return null;if {
const toolResponse = await executeTool;history.push;return toolResponse;}
history.push;return event;}
async function runDemo { /* orchestrate steps */ }
runDemo;按理说，

This demo shows that all decisions are based solely on persisted history—no temporary flags like .

Coleade vs. Goose 在审批场景中的表现对比

• Coleade:
• Goose:

User Pain Point Addressed by Goose:

• No need to keep a running process alive for hours.
• The system can recover after restarts because state is stored persistently.
• A separate worker or orchestration platform can resume task whenever events arrive.

Differentiating Design Goals: While Loop vs State Machine

• Coleade's While Loop:
• Packs complex agent behavior into a single Query Loop that handles model inference,tool execution。compression & retry logic all in one place.
• Goose's State Machine:
• Splits complex behavior into independent Operations driven by persisted conversation state.

The value of Goose emerges only when tasks span beyond a single active session—when y must survive process restarts or wait for external approvals that may take hours or days.

与思考

Coleade’s queryLoop is excellent for fast interactive scenarios where everything stays in memory and approvals happen immediately. It offers a mature orchestration of streaming model calls and tool execution with built‑in recovery mechanisms.

If your agent needs to handle longer lifecycles—external approvals that may not come back quickly,background workers that execute tools asynchronously,or deployment across multiple nodes— Goose state‑machine approach gives you durable persistence and clear separation 娱乐ween business rules and runtime orchestration.