Back to home

Compare

Comparing: Multi-Agent Failures Start with Architecture, Not Model Capability & 多 Agent 协作失败多源于架构设计,模型能力并非唯一瓶颈

AEN
Multi-AgentLangGraphToken Costs·

Multi-Agent Failures Start with Architecture, Not Model Capability

What This Is

The most common problem with multi-Agent collaboration—AI programs that can call tools, divide tasks, and work with one another—is not that the models are not intelligent enough. It is that information gets repeatedly paraphrased as it moves through the workflow, contexts become polluted, or tasks enter loops that no one notices. The article’s example: after eight back-and-forth rounds among three Agents handling requirements analysis, coding, and code review, they were still repeating the same actions, while Token consumption rose to roughly three times the original level.

The solution does not depend on using a larger model. Instead, Agents should pass references to structured artifacts rather than copies of natural-language summaries; each Agent should read only its own task and the necessary upstream materials; and the supervisor Agent should impose three kinds of limits at once: call steps, cost budget, and repeated actions. The article also provides a 120-line demo that requires no external API, offering an engineering skeleton for connecting real models later.

Industry View

We believe the article’s real value lies in pulling multi-Agent systems back from the realm of “model capability competitions” and into the category of “workflow engineering.” As model calls become increasingly affordable, enterprises can more easily add Agents. But more Agents do not automatically produce greater reliability. Traceable artifacts, context isolation, and cost control are prerequisites for deployment at scale. The article also cites data from the LangGraph team: after optimization, the same supervisor architecture performed about 50% better than its initial version, with the gap coming primarily from how information was passed.

That said, this approach is closer to a development standard and minimal skeleton than a production-ready system. The article primarily uses a deterministic “fake model” to demonstrate the flow, and it does not fully address concurrent writes, access control, artifact version conflicts, cross-service tracing, or unstable output formats from real models. The objections and reservations are equally clear: more complex multi-Agent architecture is not automatically better. For short, low-risk tasks whose results are easy to review, a single Agent with a small number of tool calls will usually cost less and expose fewer failure points.

Impact on Regular People

For enterprise IT: Requirements records, code, and review comments that were once scattered across separate AI features are beginning to form traceable collaboration chains. When enterprises procure AI projects, they need to include logging, access controls, spending limits, and failure alerts in acceptance testing, rather than comparing only the quality of model responses.

For individual professionals: This architecture will first change how software, operations, and consulting teams divide tasks. People who can design business processes, validate results, and collaborate across tools may deliver outcomes more reliably than those who know only how to write prompts.

For the consumer market: Once AI products can complete multi-step tasks consistently, services such as customer support, shopping, and travel planning will become more coherent. But errors may also be executed automatically instead of waiting to be discovered by a human. Consumers will place greater value on process records, the ability to undo actions, and whether providers clearly distinguish between “completed” and “partially successful.”

Source: juejin.cn
BZH
多 AgentLangGraphToken 成本·

多 Agent 协作失败多源于架构设计,模型能力并非唯一瓶颈

这是什么

多 Agent(能调用工具、分担任务并相互协作的 AI 程序)协作最常见的问题,不是模型不够聪明,而是信息在流程中被反复转述、上下文污染,或任务进入无人察觉的循环。文章给出的例子是:需求分析、写代码和代码审查三个 Agent 经过 8 轮来回后仍在重复动作,Token 消耗增至约 3 倍。

其解决思路并不依赖更大的模型:Agent 之间传递结构化产物的引用,而不是复制自然语言摘要;每个 Agent 只读取自己的任务和必要上游材料;主管 Agent 同时设置调用步数、费用预算和重复动作三类限制。文中还提供了一套 120 行、无需外部接口的演示代码,可作为后续接入真实模型的工程骨架。

行业怎么看

我们认为,这篇文章真正有价值的地方,是把多 Agent 从“模型能力竞赛”拉回到“工作流工程”。随着模型调用越来越便宜,企业更容易增加 Agent 数量;但数量并不能自动带来可靠性,产物可追溯、上下文隔离和成本控制才是规模化部署的前提。文中还援引 LangGraph 团队的数据:同一主管架构经过优化后,表现较初版高约 50%,差距主要来自信息传递方式。

不过,这一方案更像开发规范与最小骨架,还不能直接等同于生产可用系统。文章主要使用确定性的“假模型”演示流程,没有充分处理并发写入、权限控制、产物版本冲突、跨服务追踪,以及真实模型输出格式不稳定等问题。反对或保留意见也很明确:多 Agent 架构并非越复杂越好。对流程短、风险低、结果容易复核的任务,单个 Agent 配合少量工具调用,通常成本更低、故障面也更小。

对普通人的影响

对企业 IT:过去分散在多个 AI 功能里的需求记录、代码与审核意见,正在形成可追溯的协作链路。企业采购 AI 项目时,需要把日志、权限、费用上限和失败告警纳入验收,而不只比较模型回答质量。

对个人职场:这类架构会先改变软件、运营和咨询团队的任务拆分方式。具备业务流程设计、结果校验和跨工具协作能力的人,可能比只会写提示词的人更能稳定交付成果。

对消费市场:当 AI 产品能够持续完成多步骤任务,客服、购物、旅行规划等服务会更连贯,但错误也可能被自动执行而非等待人工发现。消费者将更看重过程记录、撤回能力,以及服务商是否明确区分“完成”与“部分成功”。

Source: juejin.cn