文 | 字母AI
K3.1 Pro发布之后，由于多位用户反馈额度烧得太快，月之暗面紧急发布公告，称“问题出在我们，我们也本应更早修复”，在补偿额度的同时，月之暗面也表示将“减少不必要的子Agent调用和输出长度”。
就在模型上线前两天，Anthropic发了一篇文章，标题为《构建Agent：何时该把子Agent折叠回主循环》（Building agents: When to fold subagents into your main agent loop），它所讲的，刚好就是关于子Agent这件事。
Anthropic的结论是，通过对比多家企业部署的Claude他们发现，一个带技能的Agent，质量反而胜过了“一群子Agent分工”，而且单任务成本和延迟往往更低。最关键的是，这个Agent所消耗的token，要远远低于一群子Agent。
子Agent这个概念，其实是模型能力不足时期，厂商为了让模型完成更多任务所打的一个“补丁”。
从任务完成率来说，多个子Agent确实要远远强于当时的单Agent。作为代价，token烧得更厉害了，模型也就变得更贵了。
Anthropic表示，随着模型变强、上下文变长以后，这个“补丁”就不再能提升能力，反而变成了一种“累赘”。
Anthropic这篇文章说了什么
Anthropic认为，对于绝大多数任务来说，你根本不需要那么多子Agent，只需要单Agent、单模型、单份对话记录、一个循环。
模型在一个loop里反复工作，每一轮可以同时发出一批工具调用，结果直接回到这个模型手里，由它决定下一步做什么。
打个比方，后厨需要做一道菜，那就得从备菜开始，再到洗菜、切菜、下锅，最后装盘上菜。
换成以前的分工，是洗菜的、切菜的、掌勺的、摆盘的各一个人，每人手里都有一个任务列表。问题就在于，每一个人的任务列表是上一个人给的，比如切菜的就是从洗菜的手里拿的。
菜端上桌发现咸了，你很难说清是哪一环的问题，有可能是掌勺的多放盐了，也有可能是菜没洗干净。
单Agent的好处在于，它不用在每一轮开头，先跟谁解释一遍“我要的是什么味道”。菜是他自己弄的，咸淡也是他自己尝过的。
文章认为，子Agent不好的理由有三个。
第一个是token复利。
任务中每多一个子Agent，就会多一层分支，成本就增多一些。多Agent系统的token消耗，约为普通对话的15倍。
关键在于这15倍不是花在“想”上，是花在“记”上。
每个子Agent都要重新建立自己的身份、规则、工具清单和相关上下文，跑完再压缩成一段摘要交回去。这就会让主Agent要读很多遍同一份材料。
第二个是信息衰减。
信息在不同的子Agent之间，每传递一次，保真度就掉一层。到后面的子Agent手里时，它拿到的可能已经和最初的需求不是一回事了。
Cognition在2025年的文章《不要构建多Agent（Don‘t Build Multi-Agents）》里写过一个场景：让两个子Agent分头做一个Flappy Bird复刻，一个做背景，一个做小鸟。结果一个做出来的背景像超级马里奥，另一个做出来的鸟根本不像游戏素材。
第三个是人格碎片。
虽然在系统内是多个子Agent之间交流，这些子Agent由于提示词不同，它们输出的风格截然不同，然而用户面对的只有一个入口。
Anthropic在文章中举例，说有一家公司，它的客户端有十几个专项子Agent，结果这个公司评价这些子Agent“很难维持稳定的品牌口径”，并且用户永远需要等最慢的那个子Agent输出完，才能得到回复。
但话要分两头说，Anthropic并不认为应该直接取缔子Agent，而是要把它们按需加载成为Skill。
每个子Agent会随身携带各种知识，就比如切菜的刀法、掌勺的火候。每到需要的时候，把这些Skill拿出来，到下一个阶段再把这些Skill放回去。
这套办法的效果是可测的。
Anthropic在文章里提到，结合配套工具，通过这种把子Agent变成Skill的方法，就可以砍掉80%以上的系统提示词。
这就让原本臃肿到会影响判断的提示词，瞬间干净简洁，因此节省出了大量的上下文空间留给任务本身。
我们真的还需要子Agent吗
子Agent在诞生之初，本就是一种工程上的妥协。
历时还身处OpenAI的卡帕西曾经发过一条推文，大致意思是说，GPT并不知道自己是有上下文限制的，让它执行任务的时候，很容易就把上下文填满，进而导致任务失败。
所以必须要通过一种手段，去拆分任务，让模型能分步完成。
2023年3月底，AutoGPT和BabyAGI做了两个实验项目。其核心都是用一个Agent把大目标拆成数个todo list，自己给自己派活，自己执行。
这批项目证明了一件事，让模型自己决定下一步，循环就能跑通。但它也暴露了一件事，如果这个循环里没有可靠的验证环节，那么错误会在循环里自我放大。
一个Agent说“我要去查这家公司的财报”，即使说查到的是错的那份，它依然会基于错的东西继续往下推，推到最后给你一份看起来很认真的胡说八道。再往后就是无限打转，错上加错。
2023年下半年到2024年，行业给出的解法是把分工做得像一家真实的公司。
比如MetaGPT，它就是把模型分成了产品经理、架构师、工程师、测试这四个岗位。
你丢一句“做一个贪吃蛇，带排行榜”，它就会按岗位一层层往下走。先由“产品经理”写出需求文档，再由“架构师”把文档翻成系统设计和接口，接着“工程师”照着接口写代码，最后“测试”出来挑毛病。由此生成完整的程序。
当时大家的想法挺朴素的，一个模型容易跑偏，那就让几个模型互相照应，这也是子Agent系统的雏形。
说白了，子Agent本质上是“用架构去模拟一台更宽、更稳、内存更大的机器”。
除了上下文窗口有限以外，Anthropic在文章中还写到，模型的可靠性也有限，链越长越容易崩。此外，模型只能单线程工作，不能同时思考。
Anthropic表示，他们在2024年做子Agent的动机，就是为了解决这三件事：上下文隔离，别让主上下文被海量中间产物污染，变相扩展了上下文窗口；不同任务给不同工具，防止链条某处崩溃时，直接影响整个任务；并行，多条独立线索同时推进。
Anthropic认为，子Agent应该算是模型的一种“补丁”。不过这些补丁，也都是要按token付钱的。
过去两年，付钱的大多是按API计量的开发者和企业，成本明码标价、可以核算。你在token上多花1毛钱，账单上就多1毛，清清楚楚。
随着C端订阅制的流行，Anthropic表示，是时候该改改了。
月度额度是一个黑箱。用户只知道“这个月还剩多少”，不知道“这一轮对话里，这些子Agent到底消耗了多少额度”。
订阅用户是按“我做了几件事”来判断额度的，可系统却是按“我调了几次模型、开了几个分支、读了几遍文件”来计算的。
因此Anthropic表示，模型性能、算力逐渐提升，这些“补丁”只有在极其复杂的任务中才能发挥作用，对于绝大多数任务来说，子Agent只会烧掉更多token。
于是Anthropic就给出了结论：子Agent带来的好处，已经抵不上它本身的成本了。
Anthropic甚至为此开源了一个名为single-agent-loop-kit的工具，专门让模型只以单Agent的形式执行任务，不再调用子Agent。
K3.1 Pro 的这张账单
月之暗面正是吃了子Agent开太多的亏。
10月24日，月之暗面的K3.1 Pro上线，性能很强。在Terminal-Bench 4.0上它拿到65.7%，超过了GPT-6 Astra的58.2%，和Claude Opus 5.5的66.4%相比，只差了0.7%。
但是社区多位用户反馈称模型额度烧得太快，随便一两个小任务就能消耗掉5小时的额度。
10月26日凌晨，月之暗面官方表示，将减少不必要的子Agent调用和输出长度。
在互联网公司打过工的人看这事一定很眼熟。
一个需求进来，产品、设计、前端、后端、测试、运营各来一位，人人都得在文档上留一句意见，不然显得没参与。到最后发现，本来一个人半天能干完的活，结果从需求到排期再到测试上线，前后大半年出去了。
虽然在公司里，它叫作组织能力。然而到了模型上，它就叫消耗token。
作为模型的用户，他只能选择模型的思考强度，没办法选择开多少个子Agent。他只想为完成任务付钱，不想为这么多子Agent“对齐、开会、做PPT”付钱。
而且用户连喊停的机会都没有。这里面有个概念，叫做失败成本前置。
正常的交易是“事做完了，钱才花掉”。子Agent是只要开了就烧额度，哪怕这个子Agent的作用几乎可以忽略不计。
公平地说，这不是K3.1 Pro独有的病。
Claude的用户也吐槽过，20美元档订阅，一个提示词就把5小时额度用完；Opus 5.5处理存量代码库，大约30分钟触发限制。都是因为Claude开了太多子Agent。
Anthropic认为，这是“长程Agent”这个品类的“结构性矛盾”。
你给它一个任务，它要查资料、读文件、写、改、再自己验一遍。它要跑几十步，每一步都得把上下文重新读一遍，跑半小时是常事。
模型越强，搜得更久，工具调得更多，自查的轮数也更多。
Anthropic给出的解决办法是，从产品侧替用户判定“哪些子Agent调用是不必要的”，进而减少子Agent，能用一个Agent解决就坚决不开子Agent。
除此以外还要让用户能决定步数上限，在单任务开始前提示一个预算区间。
眼瞅着Anthropic就要上市了，减少子Agent进而降低token消耗，也是为了讨好他们的C端用户。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Letter AI
After K3.1 Propublish, due to feedback from multiple users that quota was burning too fast, the dark side of the month was declared as an emergency publish announcement that “the problem is with us, and we should have repaired it earlier” and, while compensating quota, the dark side of the month indicated that it would “reduce the needless son Agent's call and output length”.
Just two days before the model went online, Anthropic sent an article entitled Building Agent: When to Fold Subagents into your main event Loop, which coincided with Agent.
Anthropic concludes that, by comparing Claude, which was deployed by a number of enterprises, they found that a skilled Agent had outperformed the “group of Agent division of labour” and that single-assignment costs and delays were often lower. Most importantly, this agent consumes far less token than a group of agents.
The concept of subAgent is actually a “supplier” that the manufacturer used to do more to enable the model to do.
In terms of mission completion rates, multiple sub-Agents were, indeed, much better than the single Agent at the time. At the cost of that, the token burned harder, and the model became more expensive.
Anthropic states that as the model grows stronger and context grows longer, this “compensation” no longer enhances capacity, but becomes a “burden”.
What did Anthropic say?
Anthropic believes that for most missions, you don't need so many Agents. You just need a single Agent, a single model, a single conversation, a loop.
The model works over and over again in a loop, and each round can send a batch of tool calls at the same time, and the results go straight back to the model, which determines what next step does.
For example, the cook needs to cook a dish, which starts with the preparation, then with the washing, cutting, cooking, and then the dishes.
Instead of the former division of labour, there is a list of individuals who are washing, cutting, hand-held, pallets, each with a list of tasks. The problem is that each individual's list of tasks is given by the previous person, for example, cutting vegetables from the washing hands.
It's hard for you to say which of these is the salt on the table. It's probably the salt on the palm spoon, or it's not washed.
The advantage of Agent alone is that it does not have to start with each round, explaining to whom “what I want to taste”. He made it himself, and he tasted it.
According to the article, there are three reasons why little Agent is not good.
The first is token.
Each sub-mission, Agent, has an additional layer of branch, which increases costs. The multiAgent system token consumes about 15 times more than normal dialogue.
The point is that these 15 times are spent not on thinking, but on remembering.
Each sub-unit, Agent, was to re-establish its own identity, rules, list of tools and associated context, and then compress it back into a section of summary. It's going to get the main Agent to read a lot of the same material.
The second is information decay.
The information is between different sub-Agents, and each time it is passed, the level of authenticity falls. By the time Agent was in the back, it was not the same as the original need.
Cognition wrote a scene in 2025 in Don't Build Multi-Agents: Let's split up the two children Agent to make a Flickr Bird recap, a background, a bird. And one of them was made in a context like Super Mario, and the other was made out of a bird that didn't look like a game asset.
The third one is the body fragments.
Although there are multiple sub-Agents in the system, which, because of their differences in prompt, have a different style of output, there is only one entry for users.
In an article, Anthropic cited, for example, a company whose client had more than a dozen special subsidiaries, Agent, which had the result that the company had evaluated the sub-sub-sub-sub-contractors as “difficult to maintain a stable brand calibre” and that users would always have to wait for the slowest sub-sub-sub-sub-sub-sub-contractor to be exported to get a response.
However, in two parts, Anthropic does not believe that the son Agent should be directly banned, but rather that it would make load on demand Skill.
Every child, Agent, carries all kinds of knowledge, like the cutter's knife and the hand-held fire. Whenever necessary, take out the Skills and put them back at the next stage.
The effectiveness of this approach is measurable.
In his article, Anthropic mentioned that, by combining the accompanying tools, this conversion of the sub-Agent into Skill would have cut down over 80 per cent of the system prompt.
This has allowed prompt to be so swollen as to affect judgement that it is clean and simple at once, thus saving a great deal of context space for the task itself.
Do we really need Agent?
At the beginning of his birth, son Agent was an engineering compromise.
Capaci, who was still in Openai, had sent a tweet to the effect that GPT did not know that it was subject to context, and that it was easy to fill up context when it was on mission, thus causing the mission to fail.
So there's a way to get to the split mission, so the model can be done in step.
At the end of March 2023, AutoGPT and BabyAGI carried out two pilot projects. The core is to break the big target into several todo list with an Agent, to feed itself and to execute it.
The project proved one thing, so the model itself decided next step, and the cycle ran. But it also exposes one thing, and if there's no reliable authentication in the cycle, then the error will magnify itself in the cycle.
One Agent said, "I'm going to look into the company's financials" and even if it's the wrong one, it'll continue to push it down on the basis of the wrong thing and then give you a seemingly serious piece of crap. It's going to be an infinity turn.
Between the second half of 2023 and 2024, the solution given by the industry was to make the division of labour a real company.
MetaGPT, for example, divides the model into four positions of product manager, architect, engineer and test.
You drop the phrase, "Be a snake, take a list," and it goes down one level of the post. The "product manager" first wrote the demand file, then the "architect" turned the document into a system design and interface, then the "engineer" wrote the code following the interface and finally "tested" the problem. This generates a complete program.
The idea was so simple, that a model could run away, that a few models would look after each other, and that was the prototype of the subAgent system.
To put it bluntly, sub-Agent is essentially “a much broader, more stable and larger memory machine modelled on a structure”.
In addition to the limitations of context window, Anthropic also wrote in his article that the model ' s reliability was limited and that the longer the chain, the easier it was to collapse. Moreover, the model can only work one-way and cannot be considered at the same time.
Anthropic states that their motivation to be a child of Agent in 2024 was to address these three things: the isolation of context, so that the main context would not be contaminated by a mass intermediate, with the disguised expansion of context window; the direct impact of different tasks on the whole task when different tasks are given to different tools to prevent a chain from falling off somewhere; and, in parallel, multiple independent leads to advance simultaneously.
According to Anthropic, subAgent is supposed to be a “compensation” of the model. These patches, however, are paid for by Token.
Over the past two years, most payments have been made by developers and businesses measured as API, with cost-coded and accounted for. You spend more than a cent on token and more than a cent on the bill.
With the popular C-end subscription, Anthropic said it was time to change.
Monthly quota is a black box. Users know only “how much remains this month” and do not know “how much quota was consumed by these kids in this round of conversations”.
Subscription users judge quota by “I did a few things”, but the system is calculated by “I changed several models, opened several branches, read several files”.
Therefore, Anthropic states that model performance and calculation are increasing and that these “compensation” can only work in extremely complex missions, and that for the vast majority of missions, Agent will only burn more tokens.
So Anthropic came to the conclusion that the benefits of sub-Agent were not worth its own costs.
Anthropic even opened up a tool called Single-agent-loop-kit for this purpose, which is dedicated to the use of the model only in the form of a single Agent and no longer calls the subAgent.
This bill for K3.1 Pro
The dark side of the moon is the one that ate Agent too much loss.
On 24 October, K3.1 Pro, on the dark side of the moon, had a strong performance. On Terminal-Bench 4.0, it received 65.7 per cent, which exceeds 58.2 per cent of GPT-6 Astra and represents a gap of only 0.7 per cent compared with 66.4 per cent of Claude Opus 5.5.
However, a number of community users responded that the model quota was burned too fast to consume five hours of quota for just one or two small tasks.
In the early morning hours of 26 October, the dark side of the month officially indicated that it would reduce the unnecessary call and output length of the child Agent.
Those who have worked in Internet companies must look familiar.
One needs to come in, one product, one design, one front end, one back end, one test, and one operation, and everyone has to leave a message on the document or not appear to be involved. By the end of the day, it was discovered that a person could have done half a day of work, with the result that he had left for most of the year, from demand to scheduling to testing.
Although in the company, it is called organizational capacity. But on the model, it's called consumption token.
As a model user, he can only choose the thinking strength of the model and not how many sub-Agents he can open. He only wanted to pay for his mission and not for so many children, Agent, “to align, meet, do PPT”.
And users don't even have a chance to stop. There's a concept called failure cost pre-emption.
The normal deal is “to spend the money when it's done”. quota, even if it's almost negligible.
To be fair, it's not K3.1 Pro's disease.
Claude's users have also given up, with a subscription of $20, and a prompt has run out of five hours of quota; Opus 5.5 handles the stock code library, which triggers the limit for about 30 minutes. It's because Claude drove too many kids, Agent.
According to Anthropic, this is a “structural contradiction” in the category of “long-range Agent”.
You give it a mission to look at, read, write, change, and examine itself. It has to run for dozens of steps, read context again at every step, and run for half an hour.
The stronger the model, the longer it is searched, the more tools are mobilized and the more wheels self-checked.
The solution offered by Anthropic is to determine from the side of the product which subAgent calls is not necessary for the user, and thus to reduce the subAgent, which can be solved with an Agent, and to insist that the subAgent.
In addition to this, users are allowed to decide on a maximum number of steps, indicating a budget space before a single task begins.
Anthropic is about to be listed, reducing the consumption of the child Agent and thus the consumption of the token, also to please their C-end users.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.