Kimi K3.1 Pro quota depletes too quickly; Anthropic argues for folding sub-agents back into the main loop

The Kimi quota problem: Anthropic has found a solution. Letter List Download the app for exclusive early access October 27, 2026, 11:26:17 From Beijing Updated October 27, 2026, 12:17:33

After the release of K3.1 Pro, as multiple users reported that their quota was burning up too fast, Moonshot AI urgently issued an announcement stating that "the problem is on us, and we should have fixed it earlier." While compensating quota, Moonshot AI also said it would "reduce unnecessary sub-agent calls and output length."

Just two days before the model went live, Anthropic published an article titled "Building agents: When to fold subagents into your main agent loop," and what it discussed happened to be exactly about sub-agents.

Anthropic's conclusion was that, by comparing Claude deployments across multiple enterprises, they found that a single agent with skills actually outperformed "a group of sub-agents dividing the work" in quality, and often had lower per-task cost and latency. Most critically, the tokens consumed by this agent were far lower than those consumed by a group of sub-agents.

The concept of sub-agents was actually a "patch" that vendors applied during the period when model capabilities were insufficient, in order to get models to complete more tasks.

In terms of task completion rate, multiple sub-agents were indeed far stronger than the single agent of that time. As a trade-off, tokens burned much faster, and the model became more expensive.

Anthropic said that as models become stronger and context windows grow longer, this "patch" no longer improves capability and instead becomes a "burden."

What Anthropic's article said

Anthropic believes that for the vast majority of tasks, you simply do not need that many sub-agents—you only need a single agent, a single model, a single conversation record, and one loop.

The model works repeatedly within one loop; in each round it can issue a batch of tool calls at the same time, and the results go directly back into the hands of this model, which decides what to do next.

For example, if the kitchen needs to make a dish, it has to start with preparing the ingredients, then washing them, cutting them, putting them in the pan, and finally plating and serving.

Under the old division of labor, there would be one person washing vegetables, one cutting them, one cooking, and one plating, each with their own task list. The problem is that each person's task list is given by the previous person—for example, the vegetable cutter gets theirs from the vegetable washer.

When the dish reaches the table and turns out to be too salty, it is hard to say which step caused the problem; maybe the cook added too much salt, or maybe the vegetables were not washed clean.

The advantage of a single agent is that it does not need to explain at the start of every round to someone else "what flavor I want." It made the dish itself, and it tasted the saltiness itself.

The article argues that there are three reasons sub-agents are bad.

The first is token compounding.

Each additional sub-agent in a task adds another layer of branching, and the cost increases a bit more. The token consumption of a multi-agent system is about 15 times that of an ordinary conversation.

The key point is that these 15 times are not spent on "thinking," but on "remembering."

Each sub-agent has to re-establish its own identity, rules, tool list, and relevant context, and after finishing, compress everything into a summary and hand it back. This forces the main agent to read the same material many times.

The second is information decay.

Each time information is passed between different sub-agents, its fidelity drops by one layer. By the time it reaches later sub-agents, what they receive may no longer be the same as the original requirement.

Cognition wrote about a scenario in its 2025 article "Don't Build Multi-Agents": have two sub-agents separately build a Flappy Bird clone, one doing the background and one doing the bird. As a result, one produced a background that looked like Super Mario, and the other produced a bird that did not look like game asset material at all.

The third is personality fragmentation.

Although within the system it is multiple sub-agents communicating with one another, these sub-agents have different prompts, so their output styles are completely different—yet the user faces only one entry point.

Anthropic gave an example in the article, saying that one company's client had more than a dozen specialized sub-agents, and as a result the company commented that these sub-agents "struggle to maintain a consistent brand voice," and users always had to wait for the slowest sub-agent to finish output before they could get a reply.

But there are two sides to every story: Anthropic does not believe sub-agents should be abolished outright, but rather that they should be loaded on demand as Skills.

Each sub-agent carries various kinds of knowledge with it, such as the knife skills for cutting vegetables or the heat control for cooking. Whenever needed, these Skills are taken out, and then put back at the next stage.

This approach's effect is measurable.

Anthropic mentioned in the article that, combined with supporting tools, this method of turning sub-agents into Skills can cut more than 80% of system prompts.

This instantly made the originally bloated prompt, which had grown so large it affected judgment, clean and concise, thereby freeing up a large amount of context space for the task itself.

Do we really still need sub-agents?

Sub-agents were, from the very beginning, an engineering compromise.

Karpathy, who was at OpenAI at the time, once posted a tweet to the effect that GPT does not know it has a context limit, and when you give it a task, it easily fills up the context, which then causes the task to fail.

So some means was necessary to split the task apart so the model could complete it step by step.

At the end of March 2023, AutoGPT and BabyAGI ran two experimental projects. Their core was to use one agent to break a big goal into several todo lists, assign work to itself, and execute it itself.

These projects proved one thing: if you let the model decide the next step on its own, the loop can run. But they also exposed another thing: if there is no reliable verification step in that loop, errors will amplify themselves within the loop.

An agent says "I'm going to look up this company's financial report," and even if what it finds is the wrong one, it will still keep reasoning downward based on the wrong thing, until it finally hands you a very earnest-looking pile of nonsense. After that comes endless spinning in circles, error compounding on error.

From the second half of 2023 to 2024, the industry's solution was to make the division of labor resemble a real company.

MetaGPT, for example, split the model into four roles: product manager, architect, engineer, and tester.

You throw in one line, "make a Snake game with a leaderboard," and it proceeds layer by layer according to role. First the "product manager" writes the requirements document, then the "architect" translates the document into system design and interfaces, then the "engineer" writes code against the interfaces, and finally the "tester" comes out to find faults. From this a complete program is generated.

People's thinking at the time was quite simple: if one model easily goes off track, then let several models look out for one another. This was also the prototype of the sub-agent system.

Put plainly, sub-agents are essentially "using architecture to simulate a wider, more stable machine with more memory."

Besides the limited context window, Anthropic also wrote in the article that model reliability is limited too, and the longer the chain, the more easily it collapses. In addition, models can only work single-threaded and cannot think simultaneously.

Anthropic said that their motivation for doing sub-agents in 2024 was to solve these three things: context isolation, so the main context is not polluted by massive intermediate outputs, effectively expanding the context window; giving different tasks different tools, to prevent a collapse somewhere in the chain from directly affecting the entire task; and parallelism, advancing multiple independent threads at the same time.

Anthropic believes sub-agents should count as a kind of "patch" for models. But these patches, too, must be paid for by the token.

Over the past two years, those paying have mostly been developers and enterprises billed by API, with costs clearly priced and accountable. If you spend one more tenth of a yuan on tokens, your bill goes up by one tenth of a yuan, crystal clear.

Monthly quotas are a black box. Users only know "how much is left this month," not "how much quota these sub-agents actually consumed in this round of conversation."

Subscribers judge quota by "how many things I got done," but the system calculates by "how many times I called the model, how many branches I opened, how many times I read the files."

Therefore Anthropic stated that as model performance and compute gradually improve, these "patches" only play a role in extremely complex tasks, and for the vast majority of tasks, sub-agents only burn more tokens.

So Anthropic reached its conclusion: the benefits brought by sub-agents no longer outweigh their own cost.

Anthropic even open-sourced a tool called single-agent-loop-kit for this purpose, specifically to make the model execute tasks only in single-agent form and no longer call sub-agents.

This bill for K3.1 Pro

Moonshot AI was precisely the one that suffered from opening too many sub-agents.

On October 24, Moonshot AI's K3.1 Pro launched, with strong performance. On Terminal-Bench 4.0 it scored 65.7%, surpassing GPT-6 Astra's 58.2%, and only 0.7% behind Claude Opus 5.5's 66.4%.

But multiple users in the community reported that the model burned through quota too fast, with any one or two small tasks consuming the 5-hour quota.

In the early hours of October 26, Moonshot AI officially stated that it would reduce unnecessary sub-agent calls and output length.

Anyone who has worked at an internet company will find this very familiar.

A requirement comes in, and product, design, frontend, backend, testing, and operations each send one person; everyone has to leave a comment on the document, otherwise it looks like they didn't participate. In the end you find that work one person could have finished in half a day took more than half a year from requirement to scheduling to testing and launch.

In a company, it's called organizational capability. But on a model, it's called burning tokens.

As a user of the model, he can only choose the model's thinking intensity, not how many sub-agents to open. He only wants to pay for completing the task, not for so many sub-agents to "align, hold meetings, and make PPTs."

And users don't even have a chance to call stop. There is a concept here called front-loaded failure cost.

A normal transaction is "the work is done, then the money is spent." With sub-agents, quota burns as soon as they are opened, even if that sub-agent's role is almost negligible.

To be fair, this is not a disease unique to K3.1 Pro.

Claude users have also complained that on the $20 subscription tier, a single prompt uses up the 5-hour quota; Opus 5.5 processing an existing codebase triggers the limit in about 30 minutes. All because Claude opened too many sub-agents.

Anthropic believes this is a "structural contradiction" of the "long-horizon agent" category.

You give it a task, and it has to look up information, read files, write, revise, and verify itself again. It has to run dozens of steps, and each step requires re-reading the context; running for half an hour is common.

The stronger the model, the longer it searches, the more tools it calls, and the more rounds of self-checking it does.

Anthropic's solution is to determine on the product side, on the user's behalf, "which sub-agent calls are unnecessary," thereby reducing sub-agents and resolutely not opening sub-agents when one agent can solve it.

In addition, users should be able to decide the upper limit on steps, with a budget range indicated before a single task starts.

With Anthropic about to go public, reducing sub-agents to lower token consumption is also about pleasing their consumer users.

"Special statement: The above content (including videos, pictures, and audio if any) is uploaded and published by users of Dafeng Hao, a self-media platform under Phoenix Net; this platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."