Subagents / agent swarm

  • Agents
  • Industry term
2 steps

Subagents are helper agents that a main agent hands bounded pieces of work to, each with its own context, tools, or instructions. Agent swarm is a looser label for larger or more spread-out groups of agents.

A helper works with the context its setup gives it, which may be a task summary, shared background, or a copy of earlier conversation. Separate contexts can keep one search from filling the main run's context, but the main agent still needs useful returned evidence. Files and tools remain shared unless the system isolates them.

Where you’ll meet it

You ask a research agent to compare five project-management tools for your team. It starts a subagent for each tool. The report ranks them by price, but three subagents quoted the price per user and two quoted the price for a whole team.

What to do

Before the work splits, tell the main agent what every helper must return, such as monthly price per user with a link to the pricing page, and ask it to check the answers match before comparing.

Easy to mix up

The orchestrator reads the results and chooses the next step for the whole job. Each of the subagents finishes its own piece and reports back.

Check yourself

For a second opinion on a contract summary, your agent starts a subagent using the same base model. It sees the contract and summary in a separate context. How independent is that review?

Go deeperParallel helpers suit broad research and strain on tightly linked work

Anthropic's 2025 report on its own research system described better performance on broad research evaluations and about fifteen times the tokens of ordinary chat. That is a result for one system and workload, not a general return from adding agents. Tokens are text units processed; the bill also depends on rates and other charges.

Cognition, the company behind the Devin coding agent, argued for keeping tightly coupled coding work in one agent's context in its 2025 account. Anthropic's report likewise noted that many coding tasks offer fewer independent parallel pieces than research. These accounts motivate examining task dependencies, rather than choosing an agent count from the label alone.

Practice Modern AI

Or take the learning quiz on a mix of every set