一个 Agent 解决不了的问题,未必需要一个更强的 Agent —— 也许需要三个各有专长的协作者。A problem one agent can't solve doesn't always need a stronger agent — sometimes it needs three specialists working together.
你的研究助手现在能搜、能读、能算、有记忆。下一个诱惑是让它一个人扛下整个研究 —— 但一个 researcher 解决不了的难题,未必要一个更强的 researcher,也许要拆成几个各查一摊、再加一个挑刺的。这一章讲多 agent 的三种模式,以及更重要的:什么时候根本不该上多 agent。Your research assistant can now search, read, compute, and remember. The next temptation is to make one of it carry the whole research job — but a problem one researcher can't crack doesn't need a stronger researcher; maybe it needs a few each covering one angle, plus one that pokes holes. This chapter covers three multi-agent patterns and, more importantly, when not to.
— I
一个 agent 不够,未必要更强的 agentOne Agent Falling Short Doesn't Mean You Need a Stronger One.
把一个 agent 喂得越来越胖 —— 几十个工具、超长 system —— 它反而更容易选错、更难调。有时候答案是分工,不是增强。Fattening one agent — dozens of tools, a giant system prompt — makes it more likely to choose wrong and harder to debug. Sometimes the answer is division of labor, not enhancement.多 agent 适合两种情况:上下文太大,一个 agent 的窗口装不下整个任务;或者 技能差异明显,不同子任务需要不同的工具集和提示。把一个全能 agent 拆成几个专长 agent,每个的工具更少、提示更窄、也就更准。1注 1Note 1Anthropic · 「Building Effective Agents」(2024-12-19)—— 多 agent 编排模式:routing(先分类再路由)、orchestrator-workers(协调者拆任务分给 worker);并反复强调先用最简单方案。Anthropic · 'Building Effective Agents' (2024-12-19) — multi-agent orchestration patterns: routing (classify then route), orchestrator-workers (a coordinator splits tasks among workers); and the repeated reminder to use the simplest approach first.Multi-agent suits two situations: the context is too big for one agent's window to hold the whole task; or the skills differ sharply, with different subtasks needing different tool sets and prompts. Splitting one do-everything agent into a few specialists gives each fewer tools, a narrower prompt, and thus more accuracy.1注 1Note 1Anthropic · 「Building Effective Agents」(2024-12-19)—— 多 agent 编排模式:routing(先分类再路由)、orchestrator-workers(协调者拆任务分给 worker);并反复强调先用最简单方案。Anthropic · 'Building Effective Agents' (2024-12-19) — multi-agent orchestration patterns: routing (classify then route), orchestrator-workers (a coordinator splits tasks among workers); and the repeated reminder to use the simplest approach first.判断标准是分工边界清不清楚。边界清楚(客服 vs 退款 vs 技术),拆开值得;边界模糊、子任务互相纠缠,拆开只会在 agent 之间反复传话。The test is whether the division is clean. Clean boundaries (support vs refunds vs tech) make splitting worth it; blurry boundaries with entangled subtasks just bounce messages between agents.
多 agent 不是一种东西,是几种结构。最常用的三种:路由、编排者-工人、交接。Multi-agent isn't one thing — it's a few structures. The three most common: routing, orchestrator-workers, and handoff.路由:先用一次轻量分类,把任务送给对的专长 agent。编排者-工人:一个协调者把复杂任务拆成子任务,分给各个 worker,再汇总。交接:一个 agent 把任务连同上下文交给另一个 —— OpenAI Agents SDK 把它做成了一等原语。1注 1Note 1Anthropic · 「Building Effective Agents」(2024-12-19)—— 多 agent 编排模式:routing(先分类再路由)、orchestrator-workers(协调者拆任务分给 worker);并反复强调先用最简单方案。Anthropic · 'Building Effective Agents' (2024-12-19) — multi-agent orchestration patterns: routing (classify then route), orchestrator-workers (a coordinator splits tasks among workers); and the repeated reminder to use the simplest approach first.2注 2Note 2OpenAI Agents SDK —— 把 handoff(交接)做成一等原语:一个 agent 可以把任务连同上下文交给另一个专长 agent。OpenAI Agents SDK — makes handoff a first-class primitive: one agent can pass a task, with context, to another specialist agent.Routing: a lightweight classification first sends the task to the right specialist. Orchestrator-workers: a coordinator splits a complex task into subtasks, hands them to workers, then aggregates. Handoff: one agent passes a task, with context, to another — the OpenAI Agents SDK makes this a first-class primitive.1注 1Note 1Anthropic · 「Building Effective Agents」(2024-12-19)—— 多 agent 编排模式:routing(先分类再路由)、orchestrator-workers(协调者拆任务分给 worker);并反复强调先用最简单方案。Anthropic · 'Building Effective Agents' (2024-12-19) — multi-agent orchestration patterns: routing (classify then route), orchestrator-workers (a coordinator splits tasks among workers); and the repeated reminder to use the simplest approach first.2注 2Note 2OpenAI Agents SDK —— 把 handoff(交接)做成一等原语:一个 agent 可以把任务连同上下文交给另一个专长 agent。OpenAI Agents SDK — makes handoff a first-class primitive: one agent can pass a task, with context, to another specialist agent.OpenAI 的实践指南把这些再归两大类:manager(一个中枢 agent 把别的 agent 当工具调)和 decentralized(agent 之间 handoff 交接)。一句话记法 —— manager 模式里,连线是工具调用;decentralized 模式里,连线是交接。4注 4Note 4OpenAI · 「A Practical Guide to Building Agents」(2025-04)—— 把多 agent 归两类:manager(中枢 agent 把别的 agent 当工具调)与 decentralized(agent 之间 handoff 交接);图喻:manager 的连线是工具调用,decentralized 的连线是交接。OpenAI · 'A Practical Guide to Building Agents' (2025-04) — sorts multi-agent into two families: manager (a central agent calls others as tools) and decentralized (agents hand off to one another); the graph mnemonic — in the manager pattern edges are tool calls, in the decentralized pattern edges are handoffs.OpenAI's practical guide sorts these into two families: manager (a central agent calls the others as tools) and decentralized (agents hand off to one another). One-line mnemonic — in the manager pattern the edges are tool calls; in the decentralized pattern the edges are handoffs.4注 4Note 4OpenAI · 「A Practical Guide to Building Agents」(2025-04)—— 把多 agent 归两类:manager(中枢 agent 把别的 agent 当工具调)与 decentralized(agent 之间 handoff 交接);图喻:manager 的连线是工具调用,decentralized 的连线是交接。OpenAI · 'A Practical Guide to Building Agents' (2025-04) — sorts multi-agent into two families: manager (a central agent calls others as tools) and decentralized (agents hand off to one another); the graph mnemonic — in the manager pattern edges are tool calls, in the decentralized pattern edges are handoffs.
def research(question): angles = plan_angles(question) # 把问题拆成几个角度(子问题) findings = [sub_researcher(a) for a in angles] # 每个角度一个干净窗口的子 agent report = synthesize(question, findings) # 协调者把发现汇总成报告 return critic_pass(report) # 再让一个 critic 挑刺
选哪种看任务形状:路由适合「输入种类明确、各走各的」,编排适合「一个大任务要拆开并行再合并」,交接适合「流程会在中途换专长」。别一上来就上最复杂的编排。Which one depends on the task's shape: routing for 'distinct input kinds, each going its own way,' orchestration for 'one big task to split, parallelize, and merge,' handoff for 'a flow that changes specialty midway.' Don't reach for the most complex orchestration first.
— III
多 agent 的代价,别为了多而多The Cost of Multi-Agent — Don't Multiply for Its Own Sake.
每多一个 agent,多一次上下文传递、多一份 token 账单、多一层难调的间接。多 agent 是解法,不是默认。Every extra agent adds a context handoff, a token bill, and a layer of hard-to-debug indirection. Multi-agent is a solution, not a default.Anthropic 自己的多 agent 研究系统坦白:token 用量是单 agent 的数倍,而最难的部分是协调和上下文传递 —— 信息在 agent 之间交接时会丢、会变形。3注 3Note 3Anthropic · 「How we built our multi-agent research system」(2025)—— 多 agent 的真实案例与代价:token 用量数倍于单 agent,协调与上下文传递是主要难点。Anthropic · 'How we built our multi-agent research system' (2025) — a real case and its costs: token usage several times a single agent's, with coordination and context-passing as the main difficulties.一个会在子 agent 之间反复传话的系统,慢、贵、还难复现 bug。Anthropic's own multi-agent research system is candid: token usage runs several times a single agent's, and the hardest part is coordination and context-passing — information gets lost or distorted as it's handed between agents.3注 3Note 3Anthropic · 「How we built our multi-agent research system」(2025)—— 多 agent 的真实案例与代价:token 用量数倍于单 agent,协调与上下文传递是主要难点。Anthropic · 'How we built our multi-agent research system' (2025) — a real case and its costs: token usage several times a single agent's, with coordination and context-passing as the main difficulties. A system that keeps relaying messages between sub-agents is slow, costly, and hard to reproduce bugs in.这不是 Anthropic 一家的观感。Efficient Agents(首个系统研究 agent 效率-效果权衡的工作)在 GAIA 上实测发现:很多模块是边际递减的,连记忆都「Simple Memory is Enough」—— 复杂度该按任务难度配,而不是默认拉满。5注 5Note 5Wang et al. · 「Efficient Agents: Building Effective Agents While Reducing Cost」(2025,arXiv:2508.02694)—— 首个系统研究 agent 效率-效果权衡:在 GAIA 上发现很多模块边际递减(连记忆都「Simple Memory is Enough」),复杂度应按任务难度配;其框架保留 96.7% 性能、成本几乎减半。Wang et al. · 'Efficient Agents: Building Effective Agents While Reducing Cost' (2025, arXiv:2508.02694) — the first systematic study of the agent efficiency–effectiveness trade-off: on GAIA, many modules give diminishing returns (even memory — 'Simple Memory is Enough'), and complexity should match task difficulty; its framework retains 96.7% of performance at nearly half the cost.This isn't just Anthropic's impression. Efficient Agents — the first systematic study of the efficiency–effectiveness trade-off — measured it on GAIA and found many modules give diminishing returns, even memory ('Simple Memory is Enough'): match complexity to task difficulty rather than maxing it out by default.5注 5Note 5Wang et al. · 「Efficient Agents: Building Effective Agents While Reducing Cost」(2025,arXiv:2508.02694)—— 首个系统研究 agent 效率-效果权衡:在 GAIA 上发现很多模块边际递减(连记忆都「Simple Memory is Enough」),复杂度应按任务难度配;其框架保留 96.7% 性能、成本几乎减半。Wang et al. · 'Efficient Agents: Building Effective Agents While Reducing Cost' (2025, arXiv:2508.02694) — the first systematic study of the agent efficiency–effectiveness trade-off: on GAIA, many modules give diminishing returns (even memory — 'Simple Memory is Enough'), and complexity should match task difficulty; its framework retains 96.7% of performance at nearly half the cost.所以回到第一章那条线:先用最简单的方案。1注 1Note 1Anthropic · 「Building Effective Agents」(2024-12-19)—— 多 agent 编排模式:routing(先分类再路由)、orchestrator-workers(协调者拆任务分给 worker);并反复强调先用最简单方案。Anthropic · 'Building Effective Agents' (2024-12-19) — multi-agent orchestration patterns: routing (classify then route), orchestrator-workers (a coordinator splits tasks among workers); and the repeated reminder to use the simplest approach first.多数问题,一个装备齐全的单 agent 就够了。只有当单 agent 真的撞到上下文上限或技能边界,多 agent 才开始划算 —— 而且要拆得边界清楚,不是拆得多。So back to chapter 1's line: use the simplest approach first.1注 1Note 1Anthropic · 「Building Effective Agents」(2024-12-19)—— 多 agent 编排模式:routing(先分类再路由)、orchestrator-workers(协调者拆任务分给 worker);并反复强调先用最简单方案。Anthropic · 'Building Effective Agents' (2024-12-19) — multi-agent orchestration patterns: routing (classify then route), orchestrator-workers (a coordinator splits tasks among workers); and the repeated reminder to use the simplest approach first. Most problems are handled by one well-equipped single agent. Only when a single agent truly hits a context ceiling or skill boundary does multi-agent start to pay — and then split cleanly, not split a lot.动手 · 先证明你真的需要多 agent:Hands-on · prove you actually need multi-agent first:
01
先用单 agent 顶到极限Push a single agent to its limit first
把你想拆的任务先交给一个装齐工具的单 agent,记录它具体在哪失败 —— 是上下文装不下,还是技能选错。Give the task you want to split to one well-equipped single agent first, and record exactly where it fails — context overflow, or wrong skill selection.
02
只在失败点切一刀Cut only at the failure point
按失败原因拆出最少的 agent(通常就两个),用路由或交接连起来,别上全套编排。Split out the fewest agents the failure demands (usually just two), connect them with routing or handoff, and skip full orchestration.
03
量一下代价Measure the cost
对比单 agent 和多 agent 的 token 用量与延迟。如果多 agent 没明显更好,退回单 agent —— 拆开本身不是成绩。Compare token usage and latency, single vs multi. If multi isn't clearly better, go back to single — splitting is not itself an achievement.
一个 agent 不够,
未必是它不够强,是你没拆对。.
One agent falling short
isn't always weakness — it's the wrong split..
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词,或按 Esc 回到页面Try another keyword, or press Esc to return