Central Orchestrator
A single coordinator agent receives all requests, analyzes intent, dynamically breaks down tasks, delegates to specialized worker agents, and synthesizes their outputs into a coherent response. The orchestrator decides at runtime what subtasks are needed — unlike a pipeline, the flow is not predetermined.
Structure
The orchestrator maintains conversation context, decides when to delegate versus handle directly, and can call workers in sequence or in parallel depending on task requirements.
How It Works
- Receive — orchestrator gets the user request and full context
- Analyze — determines which workers are needed and in what order
- Delegate — sends scoped subtasks to selected workers
- Collect — gathers worker outputs, handles failures and retries
- Synthesize — combines results into a unified response
The orchestrator may iterate — calling additional workers based on intermediate results.
Key Characteristics
- Dynamic task decomposition — orchestrator decides subtasks at runtime, not in advance
- Unified interface — users interact with one agent, unaware of workers
- Parallel execution — independent subtasks can run concurrently
- Single point of failure — orchestrator going down takes everything with it
- Higher cost — orchestrator reasoning + worker calls on every request
When to Use
- Tasks span multiple domains requiring different expertise
- You can't predict in advance which workers will be needed
- Users shouldn't need to know system internals
- You need to combine outputs from multiple specialists
- Clear boundaries exist between specialist domains