A single agent doing everything hits a wall — full context, blurred focus, drifting quality. Agent teams fix that by splitting the work across specialized subagents coordinated by a lead.
Every agent has one context window. Pile a big, many-part task into it and it fills up, the goals blur together, and quality slips as everything competes for attention. You've felt this when a long session starts forgetting earlier instructions.
The fix isn't a bigger prompt — it's more agents, each with a narrower job.
In an agent team, each subagent has its own context window, its own prompt, and its own tool permissions. One researches, one builds, one reviews. Because each has a clean, narrow scope, each does its part well — and you can give the reviewer different (safer) tools than the builder.
This separation is what keeps quality high as the overall job grows.
researcher -> read-only, search builder -> edit files reviewer -> read + run tests
A main agent assigns the work, lets the subagents run in parallel, and merges their results into a single answer. You talk to the lead; the lead manages the team. This composes with Dynamic Workflows — a workflow can spin up subagents as part of its plan.
One conductor, many players, one coherent result.
Claude Code's agent view lets you manage multiple sessions from one place — see which subagent is running, which is done, and where something is stuck. You step in only where you're needed instead of babysitting the whole thing.
That's how you run a team without losing the thread.
A template for defining each subagent's role, context, tools, and how the lead combines their work.