The orchestrator takes a big job and cuts it into smaller parts as needed. Each part goes to a worker built for that type of work. Workers run in parallel. The orchestrator collects the outputs and finishes the job Enables dynamic task decomposition, parallelism, and specialist processing. ## Workflow Steps 1. **Task Reception**: The orchestrator receives the high-level task or query. 2. **Task Decomposition**: The orchestrator breaks down the task into smaller, manageable subtasks based on the input. 3. **Task Delegation**: Each subtask is assigned to a dedicated worker agent specialized for that subtask. 4. **Parallel Processing**: Worker agents work independently to complete their subtasks. 5. **Result Collection**: Workers return their results to the orchestrator. 6. **Synthesis**: The orchestrator combines results, refines subtasks if needed, and repeats the process until the overall task is complete. ## Traits - Tasks shift based on the input. - Parallel work speeds things up. - Workers stay focused on one job. - Orchestrator tracks state and flow. - Good for messy or changing workflows. ## Where It Fits - Multi-agent [[AI]] systems. - Heavy data pipelines. - Complex decision flows with dependent steps. ## Implementation Notes - Orchestrator logic defines the whole system. - Workers scale independently. - Shared state holds plans and partial outputs. - Message queues improve resilience. ## References - Shares ideas with MapReduce but stays flexible. - Seen in tools like LangGraph. - Strong base for modular [[AI]] and automation setups.