How a subagent works
A subagent is a relationship rather than a type of software. A main agent, working through a job, hits a piece of work that deserves its own attention — read these competitor pages and summarise the offers, check this draft against the brand rules — and hands it to a second agent with its own instructions and its own working space. The subagent does that one thing and returns a result. The main agent carries on with the answer, not with the mess of getting it.
The important detail is separation. The subagent starts fresh, so the sprawl of a long research task stays in its own context and only the conclusion comes back. That keeps the main agent’s thread short enough to stay coherent, which is the practical reason the pattern exists. The word is recent, popularised by Anthropic’s agent tooling, and it is used loosely.
Why subagents matter
Long jobs degrade. As a conversation grows, the model has more to weigh and the earliest instructions start to lose their grip, so the tenth output is often weaker than the second. Pushing bulky work into a subagent keeps the main thread clean and the brief intact.
It is also how genuinely parallel marketing work gets done. Reviewing a set of landing pages, or profiling a list of competitors, is the same task repeated; running each as its own subagent means one bad page does not derail the rest, and the main agent simply collects what comes back.
Where subagents go wrong
The most common problem is a vague brief. A subagent cannot ask you what you meant, and it cannot see the wider job, so an instruction that made sense in your head comes back as a competent answer to a question you did not ask. Everything it needs has to travel with the request.
The second is silent loss. If a subagent fails, or returns something thin, the main agent may accept it and continue as though the task were done. Without a log of what each one was asked and what it returned, the gap is invisible in the finished work.
The third is cost. Each subagent is a fresh run of the model, so a job that fans out widely spends far more than the single conversation it resembles.
How to act on it
Use subagents for work that is bounded and self-contained, and keep the rest in one thread. Write the brief as though the reader knows nothing about the wider project, because that is exactly the situation. Ask for a defined shape of answer — a summary, a verdict, a list of findings — so the main agent can tell a real result from an empty one.
Log every call and its result, cap how many may run, and keep write permissions with the main agent rather than scattering them. Applied that way, the pattern is a sensible part of an agentic workflow and a practical building block for AI agents doing real marketing work.
