Worker Agent — AI Glossary

Reference details

Related terms

• AI Agent
• Subagent
• Orchestrator

How it works

1. The orchestrator composes a task description for the worker rather than forwarding the user's request, so the worker receives a brief it can satisfy without the surrounding problem.
2. Each worker is started with its own fresh context, its own system prompt and its own tool subset, which is why two workers in the same run can hold contradictory intermediate beliefs without interfering.
3. The worker then runs a complete loop of its own — calling tools, observing results, deciding when it is finished — rather than returning after a single turn.
4. What comes back to the orchestrator is the worker's result, not its transcript, and that discarding is the architecture's central mechanism: the orchestrator's context grows by one summary rather than by a full execution history.
5. That is also where the information loss lives, since anything the worker observed but did not report is unavailable to every later step and cannot be recovered without rerunning it.
6. The orchestrator integrates the returned results, decides whether to dispatch more work, and holds the only view of the whole task — so a failure attributable to no particular worker is an orchestrator problem by construction.

Example: Why a narrow toolset is the design

A research worker gets search and read. A code worker gets the repository and a test runner. Neither can send email. Scoping tools per worker bounds what a wrong decision can do, and it also raises tool-selection accuracy — a short, relevant list is easier to choose from than a long one.

Commonly confused with

A worker is not simply a second model call. It is a scoped agent with its own prompt, tools and permissions, delegated to by an orchestrator. Calling the same model twice with different instructions is a pipeline, and the distinction matters when you are reasoning about blast radius.

When to use it

Reach for it when:

• Sub-tasks need genuinely different tools or permissions
• Work can run in parallel and be recombined
• One context window cannot hold the whole problem

Reach for something else when:

• A single agent with the same tools would do
• You cannot attribute a bad outcome to a specific worker
• Cost matters: each worker is a full model call and they multiply

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Worker Agent — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path