Worker Agent
On a building site, the person holding the plans is rarely the person holding the drill. Multi-agent systems divide along the same line, and the one holding the drill is the worker agent.
Where it sits
A worker agent is a subordinate agent that receives a scoped task from an orchestrator, completes it with its own loop and tools, and returns a result rather than a conversation. It does not own the goal. It is told to extract the pricing table from one PDF, or to run the test suite and report the failures, and its context holds that brief plus what the brief needs — not the whole plan. That narrowness is the payoff: a focused instruction, a small tool set, and a context window that never fills with another agent's reasoning.
What the split buys
Three things a single agent cannot have at once:
- Parallelism — several workers run at the same time on independent subtasks.
- Context isolation — a worker's dead ends and raw tool output never reach the orchestrator; only its summary does.
- Specialisation — a different model, tool set or permission scope per worker: a cheap model for extraction, a stronger one for synthesis.
The orchestrator-worker graphs of LangGraph and the subagents of the Claude Agent SDK are both this shape.
Where it breaks
A worker given too little context invents the missing half of its brief and returns confident nonsense; given too much, the isolation it existed for is gone. Two workers handed overlapping briefs redo each other's work and return answers the orchestrator must now reconcile. A worker with no step limit of its own can burn the whole run's token budget on a subtask nobody is waiting for.
A worker agent is a subordinate agent that takes a scoped subtask from an orchestrator, runs its own loop and tools to completion, and returns a result rather than a dialogue. It does not own the overall goal. It matters because it keeps parallel work and noisy tool output out of the orchestrator's context: each worker holds only its own brief, so a dead end costs one subtask instead of the whole run.
On a building site, the person reading the plans and the person holding the drill are usually different people. The foreman decides what needs doing and hands out jobs; each worker gets on with their own piece and reports back when it is finished.
A worker agent is the second kind. It is a piece of AI software handed one clearly defined job — read this document and pull out the prices, run these tests and say what failed — which it does on its own and then returns an answer. It never sees the whole project, and that is deliberate: a narrow job is easier to get right.