Orchestrator-Workers Workflow

Why it matters

When the subtasks depend on the input, a fixed chain can’t plan them. This is the adaptive half of 1.6’s choice, at its smallest working size.

What to practise

Build the FlexibleOrchestrator: one call analyzes the task and returns its subtasks in XML, then a worker call per subtask receives the original task, its own approach and the shared context. Run it on marketing copy for a product, then try another product and compare the subtasks it plans. Watch the cost: N+1 calls, with workers running one at a time in this implementation. The notebook catches empty worker output; malformed XML from the orchestrator it leaves to you.

How it fits

Compare this with 1.6.1: decomposition is adaptive here because the orchestrator plans from the input. The notebook never feeds worker results back into the plan, which is the revision that 1.6.3 adds.