In plain words
An agent swarm is a loosely used term for multiple AI agents working together on a task. Each agent can handle part of the work, use tools, and pass findings to other agents or a coordinator.
A closer look
One common arrangement has a lead agent divide a goal into assignments, delegate them to specialist agents, and combine their results. Other arrangements pass control between agents. Coordination may be directed by a model, defined in code, or use both approaches. The word “swarm” alone does not specify the architecture.
Parallel agents can explore separate questions at the same time, each with its own context. This helps when a task can be split into largely independent parts. It also adds cost and coordination work: agents may repeat research, miss information, or disagree. Clear assignments, limits, and checks on the combined result matter.
In practice
To compare three products, a lead agent assigns one researcher to each product. They gather evidence simultaneously and return their findings. The lead checks gaps and conflicting claims, then writes one comparison.
A useful distinction
A swarm is not automatically decentralized, self-organizing, or better than a single agent. “Subagent” describes an agent’s delegated role; “agent swarm” describes the cooperating group. Several agents agreeing is not proof that their answer is correct.