The short answer.
Subagents are extra AI assistants that one lead assistant starts to take a piece of a job, work on it in a separate space and hand back a result; an agent team is several of them working together, sharing the work and checking each other. They help when a job splits into parts that do not depend on each other, such as reading a stack of contracts or reviewing a draft from several angles at once. They also cost more to run, are harder to follow and need clearer instructions than one assistant, so for most work in a company under 200 people, one well-built workflow in the assistant you already pay for does the job.
Most people use ChatGPT, Claude, Copilot or Gemini one conversation at a time. That is basic Excel; subagents are the advanced user’s tools, and they are easier to learn than pivot tables, because you instruct them in plain words.
What subagents and agent teams are, in plain words.
A subagent is an assistant that another assistant starts, briefs and waits on. The lead assistant breaks a job into parts, writes a short brief for each, starts a helper for each, reads what comes back and puts the answer together, like a manager who reads three researchers’ summaries rather than their notes.
- Each helper starts with a clean desk.A subagent sees its brief, not the conversation before it, so the brief has to carry everything the helper needs.
- Helpers run side by side or one after another.Parts that do not need each other run in parallel; parts that build on each other run in sequence.
- An agent team adds conversation between helpers.The helpers share a task list and message each other, so one can challenge another’s finding before the lead puts the answer together.
What ChatGPT, Claude, Copilot and Gemini offer today.
All four makers offer a version of subagents, and where each lives matters more than its name: the everyday chat product, a tool for building agents, or a coding tool or developer platform. Makers rename and move these features often; this was read on each maker’s own pages in November 2026.
- ChatGPT.OpenAI’s documentation describes subagents in ChatGPT Work and in Codex, its coding tool. ChatGPT starts specialized agents in parallel, shows their activity in the chat and gathers their results into one answer. In most cases you ask for it directly. OpenAI’s developer platform has its own version.
- Claude.In Claude Code, Anthropic’s coding tool, a subagent takes a side task in its own working space and returns only its result. Claude Code also has agent teams, marked experimental, where several sessions share a task list and message each other. On Anthropic’s developer platform, multiagent orchestration includes dynamic workflows, where Claude writes a small program that runs many agents in phases. All three are for developers and technical staff, not the everyday Claude chat.
- Copilot.In Copilot Studio, Microsoft’s tool for building agents with little or no code, a main agent can have child agents, lightweight helpers inside it, or connected agents, separate agents that more than one main agent can call. Whoever builds agents for the company sets these up there, not in the everyday Copilot chat.
- Gemini.In Gemini Enterprise, Agent Designer lets you build multi-step agents with subagents on a visual canvas, without code. Google’s Agent Development Kit (ADK) is the developer version, for teams of agents written in code.
What you gain.
Subagents give you four things a single assistant does not.
- Parallel work.Several parts of a wide job run at once, one helper per contract or competitor. The makers’ guidance says the main gain is thoroughness more than speed.
- A separate space for each part.One assistant reading a whole stack of documents starts to lose track. Each subagent reads its share apart, so the lead keeps a clear view of the whole.
- A second check.A helper that sees only the finished result reads it fresh, and catches errors the first helper talked itself into.
- Helpers set up for one job each.A helper that only checks figures against the accounting export is not distracted by instructions about tone.
What it costs you.
Subagents cost more, are harder to follow and give a job more places to go wrong. The makers say so themselves.
- More usage.Every helper does its own work, so a split job uses more of your allowance, or your bill, than one assistant doing the same job.
- Harder to follow.You read summaries, not the working, so finding which helper went wrong takes longer.
- More ways to go wrong.Context is lost at every handoff, helpers can repeat or contradict each other, and Microsoft notes that citations are not always kept when a result passes back.
- Sometimes slower.Each extra layer of coordination adds delay.
- Clear instructions or nothing.A vague brief comes back as a vague result, once per helper.
When a company under 200 people should bother.
A company under 200 people should use subagents when a job is wide, splits into parts that do not need each other, and is worth a second check before a person acts on it. A large company builds this into platforms with engineers behind them; a smaller company uses it job by job, in the assistant it already pays for. Ask these questions before you split a job.
- Does the job split into parts that do not need each other?One part per contract, client or source. If each step needs the last one’s answer, keep it in one assistant.
- Is each part big enough to be worth its own brief?Splitting a short task costs more in coordination than it saves.
- Can you write each brief in one paragraph?What to read, what to look for and what to hand back.
- Would a separate check catch errors that matter?Figures, dates and anything a client will read are worth one.
- Who reads the combined result and decides?One person, by name.
- Where does the feature live for your assistant?The everyday chat, a builder tool or a coding tool, and who in the company runs it.
- Is the extra usage worth it?A yearly review of every contract, yes. A weekly summary, rarely.
A “no” to either of the first two questions means one workflow does the job. How do we keep quality up when AI prepares the work? covers the checking that applies either way.
When one well-built workflow does the job.
One well-built workflow does the job for most repeating work in a company under 200 people: the weekly report, the follow-up letter, the month-end summary. One assistant, a written method and a person who checks is cheaper and easier to fix than a team of helpers.
The workflow also comes first. Every brief for a subagent is a small workflow, so a company that has written its methods down has the briefs ready the day a job grows wide enough to split. AI agents or workflows: which should a small company build first? makes the same case for agents in general, and How to hand a piece of work to AI sets out what each brief needs.
A brokerage splits one job and keeps the rest simple.
Take a 45-person insurance brokerage. Its everyday renewal work runs as one workflow: the assistant drafts each client’s renewal summary from the policy and the carriers’ quotes, and the account manager checks it.
Illustrative example.
Once a year the operations lead reviews every commercial policy on the book for gaps in cover. Every policy can be read on its own, so the operations lead splits the job: one subagent per policy group lists possible gaps against the same written checklist, and a separate subagent checks each list against the policy wording. The operations lead reads the combined list and decides which clients get a call. The renewal workflow stays as it is.
More from the AI you already pay for.
FullyPowered AI helps companies level up what they get from the AI they already pay for: we plan it with you, develop the workflows and deploy them, subagents included where a job earns them. The plan lives in FullyPowered Planner, and How organizations plan their use of AI sets out the method.