Following the release of K3.1 Pro, after numerous users reported that their quotas were being depleted too quickly, Moonbeam promptly issued an announcement stating, “The issue lies with us, and we should have addressed it sooner.” Alongside providing compensation for affected users, Moonbeam also indicated that it would “reduce unnecessary sub‑Agent calls and output lengths.”
Just two days before the model went live, Anthropic published an article titled “Building Agents: When to Fold Subagents into Your Main Agent Loop,” which precisely addresses the topic of subagents.
Anthropic concluded that, by comparing Claude deployments across multiple companies, they found that a single skilled agent outperformed “a swarm of sub-agents working in分工,” while also typically delivering lower per-task costs and latency. Most importantly, this single agent consumes far fewer tokens than a group of sub-agents.
The concept of a sub-agent is, in fact, a “workaround” devised by vendors during periods when model capabilities were insufficient, enabling the model to handle more tasks.
In terms of task completion rate, multiple sub-agents are indeed far superior to the single agent at the time. As a trade-off, token consumption is significantly higher, making the model more expensive.
Anthropic stated that as models grow stronger and context lengths increase, this “patch” no longer enhances performance and instead becomes a “burden.”
01
What does this article by Anthropic say?
Anthropic believes that for the vast majority of tasks, you don’t need so many sub-agents at all—just a single agent, a single model, a single conversation log, and one loop.
The model operates repeatedly within a single loop; in each iteration, it can simultaneously issue a batch of tool calls, with the results returned directly to the model, which then decides what to do next.
To give an example, when the kitchen needs to prepare a dish, it starts with prepping the ingredients, then moves on to washing and chopping them, followed by cooking, and finally plating and serving.
Under the previous division of labor, there was one person each for washing vegetables, chopping vegetables, cooking, and plating, with each person carrying a task list. The problem was that each person’s task list was handed down from the previous worker—for example, the person chopping vegetables would get their tasks from the person who washed the vegetables.
When the dish is served and you find it too salty, it’s hard to pinpoint which step went wrong—perhaps the chef added too much salt, or maybe the ingredients weren’t washed thoroughly.
The advantage of a single agent is that it doesn’t have to explain at the start of each round, “What flavor am I looking for?” The dish is prepared by the agent itself, and it has also tasted the seasoning itself.
Author’s statement: This image was generated by AI.
The article argues that there are three reasons why sub-agents are problematic.
The first is token compounding.
For each additional sub-Agent in a task, an extra layer of branching is added, which increases the cost somewhat. The token consumption of a multi-Agent system is approximately 15 times that of a standard conversation.
The key is that this 15x isn’t spent on “thinking,” but on “remembering.”
Each sub-Agent must re-establish its own identity, rules, tool inventory, and relevant context, then compress the results into a summary before sending it back. This forces the main Agent to read the same material multiple times.
The second is information decay.
Each time information is passed between different sub-agents, its fidelity diminishes with every hop. By the time it reaches the later sub-agents, what they receive may no longer match the original requirement.
In its 2025 article “Don’t Build Multi-Agents,” Cognition described a scenario: two sub-agents were tasked with independently recreating Flappy Bird—one handling the background, the other the bird. As a result, one produced a background that resembled Super Mario, while the other created a bird that bore no resemblance to the game’s original assets.
The third is Personality Fragments.
Although communication occurs among multiple sub-agents within the system, these sub-agents produce outputs with markedly different styles due to their distinct prompting instructions; nevertheless, users interact with only a single entry point.
Anthropic cited an example in its article, noting that one company had over a dozen specialized sub‑agents on its client side. The company found these sub‑agents “difficult to maintain consistent brand messaging,” and users always had to wait for the slowest sub‑agent to finish its output before receiving a response.
However, it’s important to note that Anthropic doesn’t believe sub-agents should be outright banned; instead, they propose loading them on demand as Skills.
Each sub-Agent carries a variety of knowledge with it, such as knife skills for chopping vegetables and the proper heat control for cooking. Whenever needed, these Skills are deployed; once the next stage is reached, they are put back.
The effectiveness of this approach is measurable.
Anthropic mentions in its paper that, when combined with supporting tools, this approach of turning sub‑agents into skills can reduce system prompts by more than 80%.
This makes the originally cumbersome prompt, which could hinder judgment, instantly clean and concise, thereby freeing up substantial contextual space for the task itself.
02
Do we really still need sub-agents?
From the very beginning, sub-agents were essentially an engineering compromise.
At the time, Kapasi, who was still at OpenAI, once posted a tweet to the effect that GPT is unaware of its contextual limitations; when tasked with an assignment, it can easily exhaust its context, leading to task failure.
Therefore, it is necessary to use a method to break down the task so that the model can complete it step by step.
At the end of March 2023, AutoGPT and BabyAGI conducted two experimental projects. At their core, both use an agent to break down a large goal into several to-do lists, assign tasks to itself, and execute them independently.
This batch of projects proves one thing: when the model is allowed to decide its next step on its own, the loop can run smoothly. However, it also reveals another issue: if the loop lacks a reliable validation stage, errors will propagate and amplify within the loop.
An agent says, “I’m going to check this company’s financial report,” and even if it ends up finding the wrong one, it will still proceed based on that incorrect information, ultimately delivering a seemingly thorough but entirely fabricated nonsense. After that, it just spins in circles, compounding error upon error.
From the second half of 2023 through 2024, the industry’s proposed solution is to organize division of labor as if it were a real company.
For example, MetaGPT divides the model into four roles: product manager, architect, engineer, and tester.
You just toss in the phrase “Build a Snake game with a leaderboard,” and it cascades down through the roles. First, the “Product Manager” drafts the requirements document; then, the “Architect” translates that document into system design and APIs; next, the “Engineer” implements the code based on those APIs; and finally, the “Tester” identifies and reports any issues. This process yields a fully functional application.
At the time, the idea was quite simple: if a single model tended to drift off course, why not have several models cross-check each other? This was essentially the early prototype of a sub-agent system.
To put it simply, a sub-agent is essentially “using an architecture to emulate a machine that’s wider, more stable, and has more memory.”
In addition to its limited context window, Anthropic also notes in the paper that the model’s reliability is limited and that it becomes increasingly prone to failure as the chain grows longer. Furthermore, the model can only operate in a single thread and cannot think simultaneously.
Anthropic stated that their motivation for developing sub‑agents in 2024 was to address three key issues: context isolation, preventing the main context from being polluted by a large volume of intermediate outputs and thereby effectively expanding the context window; assigning distinct tools to different tasks to ensure that a failure at one point in the chain does not directly impact the entire task; and parallelism, enabling multiple independent threads to advance simultaneously.
Anthropic believes that sub-agents should be regarded as a kind of “patch” for the model. However, these patches also incur costs billed by token.
Over the past two years, most of the paying users have been developers and enterprises billed by API usage, with costs clearly priced and easily calculable. If you spend an extra 10 cents on a token, that amount shows up on your bill—clear and straightforward.
With the growing popularity of consumer-facing subscription models, Anthropic says it’s time for a change.
Author’s statement: This image was generated by AI.
The monthly quota is a black box. Users only know “how much remains for this month,” but they don’t know “how much quota these sub‑agents have actually consumed in this round of the conversation.”
Subscribers’ quotas are determined based on “how many tasks I’ve completed,” but the system calculates them according to “how many times I’ve tuned the model, opened how many branches, and read how many files.”
Therefore, Anthropic states that as model performance and computational power continue to improve, these “patches” only prove effective in extremely complex tasks; for the vast majority of tasks, sub-agents merely end up consuming more tokens.
Thus, Anthropic concluded that the benefits brought by sub-agents no longer outweigh their own costs.
Anthropic even open-sourced a tool called single-agent-loop-kit for this purpose, specifically designed to enable models to perform tasks as single agents without invoking sub-agents.
03
This invoice for the K3.1 Pro
The Dark Side of the Moon suffered greatly from having too many child agents.
On October 24, Moonlight’s K3.1 Pro was launched, boasting strong performance. On Terminal-Bench 4.0, it scored 65.7%, surpassing GPT-6 Astra’s 58.2% and falling just 0.7% short of Claude Opus 5.5’s 66.4%.
However, many users in the community have reported that the model’s quota is being depleted too quickly—just one or two small tasks can consume an entire 5-hour quota.
In the early hours of October 26, Moonlight’s official team announced that it would reduce unnecessary sub-agent calls and output lengths.
Anyone who’s worked at an internet company will surely find this very familiar.
When a new requirement comes in, one person from each team—product, design, frontend, backend, testing, and operations—has to chime in with a comment on the document; otherwise, it looks like they weren’t involved. In the end, what could have been done by one person in half a day ends up taking nearly six months—from requirements gathering through scheduling to testing and deployment.
Although in the company it’s called organizational capability, on the model it’s referred to as token consumption.
As a user of the model, he can only choose the model’s thinking intensity and has no way to decide how many sub‑agents to activate. He just wants to pay for task completion, not for all those sub‑agents to “align, hold meetings, and create PPTs.”
Moreover, users don’t even have a chance to call for a halt. There’s a concept here called “front-loading the cost of failure.”
In a normal transaction, “the money is only spent after the task is completed.” By contrast, a sub-agent burns its quota as soon as it’s activated, even if its function is almost negligible.
To be fair, this isn’t a problem unique to the K3.1 Pro.
Claude users have also complained that with the $20 subscription plan, a single prompt can exhaust the 5-hour quota; and with Opus 5.5 processing an existing codebase, the limit is triggered in about 30 minutes. All of this is due to Claude launching too many sub‑Agents.
Anthropic believes this is a “structural contradiction” inherent in the “long-range Agent” category.
You give it a task, and it has to research information, read documents, write, make revisions, and then verify everything on its own. It needs to execute dozens of steps, with each step requiring it to re-read the context—running for half an hour is quite common.
The stronger the model, the longer the search takes, the more tools are invoked, and the greater the number of self-check iterations.
Author’s statement: This image was generated by AI.
Anthropic’s solution is to have the product itself determine for users “which sub‑agent calls are unnecessary,” thereby reducing the number of sub‑agents—when a single agent can handle the task, sub‑agents are firmly avoided.
In addition, users should be able to set a step-count limit and receive a budget range prompt before starting a single task.
With Anthropic poised to go public, reducing sub‑agents to lower token consumption is also a move to win over their end users.