agent 的规划Planning for agents

规划就是先决定步骤序列，再逐步执行。 与之相对的是边做边定：模型每一轮只看当前局面决定下一步。 两种范式没有高下，只有适不适合—— 判断依据是：步骤数能不能预先说清。

Planning means deciding the step sequence first, then executing step by step. The opposite is deciding as you go: each round the model looks only at the current situation and picks the next step. The two paradigms have no superiority — only fit — and the criterion is: can the number of steps be stated in advance.

Anthropic 把「流程可预定义」的系统叫 workflow。 Anthropic 官方原文："Workflows are systems where LLMs and tools are orchestrated through predefined code paths."

Anthropic 官方原文：与之相对的 agent 是 "LLMs dynamically direct their own processes and tool usage"。 对应的三种规划模式：

提示链（prompt chaining）：把任务拆成固定子任务序列，上一步输出喂下一步。 Anthropic 官方原文："Prompt chaining decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one." 代价是延迟，换来的是每一步变简单、更可控。 官方建议在中间步骤加程序化检查（gate），确认流程还在轨道上。

路由（routing）：先分类，再进不同分支。 适合「输入有明确类别、每类处理方式不同」的场景。

编排者-工作者（orchestrator-workers）：中央 LLM 动态拆任务、派给 worker、再合成结果。 适合子任务无法预先枚举的场合——比如改代码， 要动几个文件、每个文件怎么改，得看了任务才知道。 官方点出它和并行化的关键区别： Anthropic 官方原文："the key difference from parallelization is its flexibility— subtasks aren't pre-defined, but determined by the orchestrator based on the specific input."

Anthropic calls systems with a predefinable flow workflows. Anthropic official text: "Workflows are systems where LLMs and tools are orchestrated through predefined code paths."

Anthropic official text: the contrasting agent is "LLMs dynamically direct their own processes and tool usage". The three planning patterns that follow:

Prompt chaining: decompose the task into a fixed sequence of subtasks, each step's output feeding the next. Anthropic official text: "Prompt chaining decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one." The price is latency; what you gain is each step becoming simpler and more controllable. The official advice adds programmatic checks (gates) at intermediate steps to confirm the flow is still on track.

Routing: classify first, then branch. Fits "inputs have clear categories, each category handled differently" scenarios.

Orchestrator-workers: a central LLM dynamically decomposes tasks, dispatches workers, then synthesizes results. Fits cases where subtasks cannot be enumerated in advance — like changing code: how many files to touch and how to change each one only becomes clear after seeing the task. The official docs draw the key difference from parallelization: Anthropic official text: "the key difference from parallelization is its flexibility— subtasks aren't pre-defined, but determined by the orchestrator based on the specific input."

什么时候必须放弃固定流程？官方给了判据：

"Agents can be used for open-ended problems where it's difficult or impossible to predict the required number of steps, and where you can't hardcode a fixed path."

步骤数说不清，就别画流程图。 但边做边定有三个老问题：跑偏、循环、半路而废。 对应的工程手段是给执行加反馈环。

Anthropic 的评估者-优化者（evaluator-optimizer）是一个 LLM 生成、 另一个 LLM 评估并反馈，循环打磨。 官方给的两个适配信号： 反馈能实质改善输出；模型自己给得出这个反馈。

Reflexion 论文（arXiv 2303.11366）把反馈做成了体系。 agent 用语言复盘任务反馈，把反思文本存进 episodic memory buffer， 下次试同类任务时先读教训再动手。 论文报告 HumanEval 编码基准上 91% pass@1， 超过当时 GPT-4 的 80%。 注意这是论文在特定基准上的结果，不是通用承诺。

When must you abandon a fixed flow? The official criterion:

"Agents can be used for open-ended problems where it's difficult or impossible to predict the required number of steps, and where you can't hardcode a fixed path."

If the step count cannot be stated, don't draw a flowchart. But deciding-as-you-go has three old problems: drifting off course, looping, quitting halfway. The matching engineering answer is adding a feedback loop to execution.

Anthropic's evaluator-optimizer is one LLM generating, another LLM evaluating and giving feedback, polishing in a loop. The two fit signals the official docs give: feedback can materially improve the output; the model can produce that feedback itself.

The Reflexion paper (arXiv 2303.11366) systematizes feedback. The agent verbally reviews task feedback, stores the reflection text in an episodic memory buffer, and reads the lessons before attempting a similar task next time. The paper reports 91% pass@1 on the HumanEval coding benchmark, beating GPT-4's 80% at the time. Note these are paper results on a specific benchmark, not a general promise.

让流程可中断、可恢复，是规划里最实用的工程投资。 LangGraph 的官方文档把 human-in-the-loop 和 fault tolerance、time travel 并列为 checkpointer 的用途： 状态有快照，才能在某一步停下等人批，再从原点继续。

关键步骤交给流程，危险步骤还给人。 判断哪步危险：不可逆、要花钱、会发出去。

Making the flow interruptible and resumable is the most practical engineering investment in planning. LangGraph's official docs list human-in-the-loop alongside fault tolerance and time travel as checkpointer uses: only with state snapshots can you pause at a step for human approval and continue from the origin.

Key steps go to the flow, dangerous steps stay with humans. How to judge a dangerous step: irreversible, costs money, or goes outbound.

1. 计划过期：环境变了，旧计划照跑到底。 2. 过度规划：为五分钟能做的小事先写五千字计划。 3. 假反馈：评估者只会说「很好」，优化循环原地打转。 4. 无恢复点：跑到一半崩了，只能从头再来。

通用原则来自 Anthropic 官方原文： "Success in the LLM space isn't about building the most sophisticated system. It's about building the right system for your needs." 从最简单的提示词开始，配上扎实的评测， 简单方案不够用了，再加多步系统。

1. Stale plans: the environment changed, the old plan runs to the end anyway. 2. Over-planning: writing a five-thousand-word plan for something doable in five minutes. 3. Fake feedback: the evaluator only says "looks good", the optimization loop spins in place. 4. No recovery point: it crashes halfway and you can only start over.

The general principle, from Anthropic official text: "Success in the LLM space isn't about building the most sophisticated system. It's about building the right system for your needs." Start from the simplest prompt with solid evaluation; add multi-step systems only when the simple approach stops being enough.