Subagents / agent swarm

A pattern where a main agent delegates distinct subtasks to specialized child agents, each working in its own context. A "swarm" is the same idea at larger scale, with many agents coordinating or running in parallel.

When a task is too large or too varied for one agent to handle well, the main agent can spin up subagents, each focused on a specific piece. For example, a code-review agent might delegate security analysis to one subagent, performance review to another, and style checks to a third, then collect their results. Each subagent has its own context window (the information it can see), which keeps them focused and prevents one subtask from crowding out another. A "swarm" is the same concept at a larger scale: dozens or hundreds of agents working in parallel on different parts of a problem.

Builder example

Subagents give you speed (parallel execution) and focus (each agent sees only what it needs). The cost is coordination complexity: subagents can produce conflicting results, duplicate work, or fail silently while the main agent moves on. Every subagent also adds to your token costs and makes the overall system harder to debug, since you need to trace through multiple agents to reconstruct what happened.

Common confusion: Adding more agents does not make a system smarter. Poorly scoped subagents multiply confusion and cost. Each subagent needs a clearly bounded task, a defined output format, and an owner responsible for its results.