Automation and AI

Subagent

Also called Sub-agent, worker agent

An AI agent that another agent calls to handle one bounded task and report the result back.

Quick facts: Subagent

Category
Automation and AI
Also called
Sub-agent, worker agent
Level
Advanced
Affects
Output quality on long jobs, running cost, traceability
Where to see it
n8n agent nodes, custom builds on model provider APIs, agent development frameworks
In this article4
  1. How a subagent works
  2. Why subagents matter
  3. Where subagents go wrong
  4. How to act on it

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.

Do and do not

Do

  • Write each brief as if no context exists
  • Ask for a defined shape of answer
  • Log what each subagent was asked and returned

Do not

  • Split work that fits in one thread
  • Assume a returned result means the task succeeded
  • Hand write permissions to every subagent

Questions people ask about this

What is the difference between a subagent and a multi-agent system?

A subagent describes the relationship between two agents: one calls the other for a bounded task and gets a result back. A multi-agent system describes the whole arrangement of several agents working on one job. A system built from subagent calls is one common shape a multi-agent system takes, but the words answer different questions.

Why not just let one agent do everything?

Because long jobs drift. As a single conversation grows, the model has more to weigh and the original brief loses its grip, so later output is often weaker than early output. Handing bulky, self-contained work to a subagent keeps the main thread short and the instructions clear, at the cost of another model run.

Do subagents make the work more expensive?

Yes, because each one is a separate run of the model rather than a continuation of the same conversation. A job that fans out across many subagents can cost considerably more than it looks. Cap how many may run at once, and reserve the pattern for tasks bulky enough to justify the extra spend.

Related terms

Found this useful?

Share it, or ask an AI to summarise it

Back to the glossary

Knowing the term is the easy part

Applying it to your own site and budget is the work. Book a call and I will tell you what actually applies to you.