A narrowly scoped AI agent that a main agent hands a specific piece of work to, with its own instructions and its own clean context. For agencies, it is the closest thing AI has to delegating a task to a specialist rather than asking a generalist to do everything.
Also known as sub-agent, child agent, worker agent
A subagent is an agent spawned by another agent to handle one bounded job and report back. The parent agent keeps the plan and the overall goal. The subagent gets a tight brief, its own fresh context window, and usually a restricted set of tools, then returns a result rather than a transcript of everything it tried.
The reason this structure caught on through 2026 is practical rather than philosophical. A single agent working a long task accumulates dead ends, tool output, and abandoned reasoning in its context, and its judgment degrades as that pile grows. Handing a chunk of work to a subagent contains that mess inside the subagent, so the parent only ever sees the answer.
Agencies already run this pattern with people, which makes it one of the easier AI concepts to get a team to actually adopt.
It maps onto how studios already work. A lead holds the brief and pulls in a researcher, a copywriter, and a proofer, each briefed narrowly. Subagent architecture is that org chart written as software, which means the people designing the workflow already have the instincts for it.
It keeps long jobs from getting dumber. A competitive audit that touches forty sources will overwhelm a single agent. Split across subagents, each one reads a handful of sources and returns a summary, and the parent stays sharp because it never sees the raw forty.
It contains blast radius. You can give a research subagent read-only web access and no ability to write files or send anything. The parent stays capable, the risky steps stay boxed in.
An agency builds a new business research workflow. The parent agent takes a target company name and produces a one-page point of view. It spawns four subagents in parallel: one reads the company’s last two years of press coverage, one pulls their current creative from ad libraries and describes the pattern, one summarizes their leadership changes, and one checks what AI answer engines currently say about them. Each returns roughly two hundred words. The parent never touches the underlying pages, it just receives four tight summaries and writes the point of view from those. What used to be a strategist’s afternoon becomes a review of four summaries and an argument about what they mean.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.