You have been asking one chat window to do five jobs at once. Research the topic, draft the post, check the facts, rewrite the hook, and remember your pricing while it does the other four.
There is a name for the alternative, and it is subagents. The pages explaining it were written for people who write software: Anthropic’s own course for developers, Microsoft’s Visual Studio Code documentation, LangChain’s framework docs, OpenAI’s ChatGPT learning site, a Reddit thread in a coding community and a run of engineering blogs. Not one of them is written for a person running a business alone.
I watched that gap in person this fall, at an AI festival in Berlin. A large enterprise software company had a slot on the main stage, and the speaker talked about harnesses and about wiring forty-two separate data sources into one system. It was a good talk. It was also a careful description of a problem neither you nor I have. A one-person business has a handful of sources, not forty-two, and the useful question is never how to orchestrate them. It is what you have actually written down about any of them.
That gap is worth closing, because this fall made the same point from the other direction. On 22 September 2026 Anthropic released Claude Opus 5.5 and OpenAI answered with two cheaper GPT-6 models, one of them priced at ten cents per million input tokens against four dollars for Opus 5.5, as reported by SiliconANGLE. Those are developer prices rather than the monthly subscription you pay, and the direction is the part that matters, because the models got better and the cheap end got cheaper in the same week. Whatever is limiting your content this month, it is not the intelligence of the model and it is not the bill.
What are subagents?
A subagent is a narrow helper that an AI assistant hands one job to, working in its own separate space, and returning only a short result to the main conversation. The category it belongs to is multi-agent systems, which is the practice of splitting work that one assistant used to do alone across several assistants with narrower briefs. The mechanism is separation: the helper gets its own instructions and its own working memory rather than inheriting the whole conversation, so the job it does is not diluted by everything else you have been discussing. The result is that the main conversation stays about the thing you came for, and the narrow job gets done by something that was told about that job and nothing else.
Anthropic’s own course puts it in one line: subagents are “specialized assistants that Claude Code can delegate tasks to.” Microsoft’s Visual Studio Code documentation, updated 21 October 2026, describes the same shape functionally: “A subagent works in its own context and returns a focused result to the main agent.” Philipp Schmid, writing on 15 September 2025, defines it tighter still: “A subagent is a specialized Agent that is purpose-built for a single, well-defined task.”
Three separations, because the word is busy. In AI, a subagent is the narrow helper described above. In real estate and insurance, a sub-agent is a person who works on behalf of another agent, with its own contracts and its own body of law, and it is the older meaning by a long way. In UK civil engineering, a sub-agent is a site management job title with a salary band attached.
The grammar does the sorting, which is lucky. People asking about the software use the plural, and people asking about brokerage use the singular. If you arrived here typing the plural, the AI meaning is the one you wanted.
Why did developers start splitting one assistant into several?
They ran out of room, and then they noticed the quality problem underneath the room problem. Anthropic’s course describes the first half in two sentences. The first is “Every time you chat with Claude Code, you’re adding to the main context window”, and the second is “That space is finite, and once it fills up, Claude starts losing track of earlier parts of the conversation.”
Schmid names the second half, which is the more interesting one. “When a single, big and complex agent handles many tasks, its context window, number of tools, can become cluttered and less reliable.” The problem is not only that a long conversation runs out of space. It is that an assistant asked to do six unrelated things in one thread does all six of them slightly worse than an assistant asked to do one.
The fix the developer world settled on is to give the narrow job its own instructions and its own space. LangChain’s documentation describes the arrangement as a main agent, often called a supervisor, that calls subagents as tools, and notes that this is what prevents “context bloat in the main conversation”. Anthropic’s course describes what comes back when the helper finishes: “only a summary comes back to your main conversation.”
What is the difference between a subagent and a skill?
A subagent is a helper you hand a job to, and a skill is an instruction sheet that changes how a job gets done. That is the whole distinction, and it is worth stating because enough people now search the two side by side that Google suggests the comparison on its own.
A skill does not get its own space. It is guidance that travels with the assistant you are already talking to, so it shapes the answer inside the conversation you are in. A subagent does get its own space, which is the reason to use one, and the reason it costs more. Microsoft’s documentation is blunt about what that separation takes away: “A Local subagent doesn’t inherit the main conversation history”, and “Each Local invocation is stateless: the main agent can’t send follow-up messages to the same subagent.” In plain terms, the helper starts from nothing every time, and you cannot go back and ask it a follow-up question.
Both approaches are fair choices and they answer different questions. If the problem is that the assistant does not know how you want a thing done, the answer is written instructions. If the problem is that you have asked one thread to do four unrelated jobs at once, the answer is separation. Most of the disappointment people report with AI output is the first problem, not the second, which is worth knowing before anyone goes shopping for architecture.
What does this mean if you run your business alone?
It means the argument the developer world just had is your argument, and you can have it without writing a line of code.
Here is the shape of it. A solopreneur opens one chat and asks it to research the topic, draft the post, fix the voice, cut it down for Instagram and write the captions. That is five jobs in one thread with one brief between them, which is precisely the arrangement engineering teams stopped using this year. They did not fix it with a better prompt and they did not fix it by waiting for a smarter model. They fixed it by writing down what each narrow job actually needs and handing that job its own instructions.
You can do the same thing with nothing more exotic than separate conversations and separate written briefs. One conversation only researches, and its brief says what a good source looks like to you. A second only drafts, with your positioning and your offer in front of it, and a third only edits for voice, working from your voice rules written down rather than described from memory. The separation is not the clever part. The written brief is the clever part, and it is the part almost nobody has, which is why the output comes back generic no matter which model is answering.
The most impressive AI setup I have seen this year belongs to a colleague in Egypt, and the impressive part was the order he built it in. The knowledge base came first. The tools were connected to it after that. The skills came last, once there was something for them to work from. None of that is architecture, and it is the reverse of how most people start, which is with whichever piece sounds cleverest.
That is the structural version of the fix, and it survives every price cut and every new model release, because it is about what you hand over rather than what you hand it to. Business Brain, the seven-day course that builds you a knowledge base for AI, is built around that one move. It is €67 one time, four modules across 22 lessons, about an hour a day for a week, in English, and it needs no code beyond one backup step in GitHub Desktop.
When is a subagent the wrong answer?
It is the wrong answer most of the time, and almost always if you run a business alone. The three limits below all come from the people who build these systems, quoted from their own documentation.
The first limit is cost, and the figure is Anthropic’s own. In its engineering write-up of 13 June 2025 on building a multi-agent research system, Anthropic states that “agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats.” The same piece sets the condition plainly: “For economic viability, multi-agent systems require tasks where the value of the task is high enough to pay for the increased performance.” Fifteen times the tokens to write an Instagram caption is not a trade anybody should want.
The second limit is that separation cuts both ways. Anthropic names it directly: “some domains that require all agents to share the same context or involve many dependencies between agents are not a good fit for multi-agent systems today.” Work where every step depends on the exact detail of the step before it is the wrong shape for handing pieces to helpers who cannot see each other.
The third limit is practical and it is the one that matters most here. Subagents are a developer feature. They live in coding tools and agent frameworks, they are configured in files, and Microsoft’s documentation even specifies that nesting them is “limited to a maximum depth of five.” If you work in a normal chat window, you very likely cannot switch one on, and you do not need to. Schmid, who is enthusiastic about the pattern, still writes that “reliability is still a challenge for agentic systems”, even with a subagent architecture in place.
What transfers to a one-person business is not the feature. It is the finding underneath it, which is that a narrow job with its own written brief beats a wide job with none. You can act on that this afternoon, with the tools already open on your screen.
The short version
- A subagent is a narrow AI helper that works in its own separate space and returns a short result to the main conversation.
- Anthropic’s Claude Academy course defines subagents as “specialized assistants that Claude Code can delegate tasks to.”
- Microsoft’s Visual Studio Code documentation, updated 21 October 2026, states that a subagent “works in its own context and returns a focused result to the main agent.”
- Anthropic’s engineering write-up of 13 June 2025 reports that multi-agent systems use about 15 times more tokens than chat interactions, and that agents use about 4 times more.
- Microsoft’s Visual Studio Code documentation states that subagent nesting is limited to a maximum depth of five.
- In real estate and insurance a sub-agent is a person acting on behalf of another agent, which is an older and unrelated meaning of the same word.
- Business Brain is a seven-day course that builds a knowledge base for AI, priced at €67 one time, four modules across 22 lessons, in English.
What to do this week
Find out what your AI actually knows about your business before you decide it needs a better architecture. The Business Brain AI Audit is free and it is the honest first step: take the audit.
If you already know the answer is “almost nothing”, Business Brain is the seven-day version of the fix.
Frequently asked questions
What are subagents in Claude and ChatGPT?
They are narrow helpers that the coding and agent versions of these tools can hand a single job to, each working in its own space and reporting back briefly. Anthropic documents them in Claude Code and in its Claude Academy course. The ordinary chat interfaces most people use, in Claude and in ChatGPT alike, do not generally expose them.
Are subagents worth it?
It depends entirely on how valuable the job is. Anthropic’s own guidance is that multi-agent systems need tasks where the value is high enough to pay for the extra performance, and its measured figure is about fifteen times the tokens of a chat. For a research job that would take you a day, that math can work. For a caption, it does not.
Do subagents use more tokens?
Yes, considerably. Anthropic reports that agents typically use about four times more tokens than chat interactions and that multi-agent systems use about fifteen times more. The separation that makes each helper focused is also what makes the whole arrangement expensive.
When should you use a subagent?
When one job is genuinely separate from the rest of the conversation and would otherwise crowd it. LangChain’s documentation suggests them for several distinct domains, for work that does not need to talk to the user directly, and where you want one place controlling the sequence.
Can subagents talk to each other?
Generally not. In the common arrangement a main agent calls each helper and collects the results, and the helpers do not see one another. Microsoft’s documentation adds that a local subagent cannot invoke further subagents by default.
How do subagents work?
The main assistant writes a short description of the job and passes it to a helper that has its own instructions. The helper works through it separately and sends back a result rather than a transcript. Anthropic’s course describes the return as a summary coming back to the main conversation.
What is the difference between a subagent and an agent?
An agent decides what to do and does it. A subagent is an agent working under another one, given a single well-defined task rather than the whole problem. Schmid’s definition is the tightest available: a subagent is “purpose-built for a single, well-defined task.”
Is a subagent in AI the same as a sub-agent in real estate?
No, and the two meanings have nothing to do with each other. In real estate and insurance a sub-agent is a person who acts on behalf of another agent under a contract. In AI it is a piece of software given one narrow job by another piece of software.
