Orchestrator–workers

All steps

Normal run

1. Request arrives. A single agent owns the task end to end. At this point its context holds only the customer's message.
2. Plan. The orchestrator decomposes the request into three independent subtasks. Planning is where the large model earns its cost.
3. Dispatch in parallel. Each subtask goes to a worker with its own clean context. The three workers start at the same moment.
4. Workers execute. Workers call tools and read documents in isolation. None of that material enters the orchestrator's context.
5. Condensed results return. Each worker returns one structured fact instead of its full transcript.
6. Synthesise. The orchestrator combines three short facts into a single decision.
7. Respond. Refund approved. Wall-clock time is bounded by the slowest worker, not the sum of all three.

Outcome: correct answer. The large model only planned and decided. Heavy reading ran in parallel, in isolated contexts, and came back as compact facts.

Failure mode

1. Request arrives. Identical start to the normal run.
2. Plan. The orchestrator produces the same three subtasks.
3. Dispatch in parallel. Workers start in parallel, as before.
4. Workers execute. Nothing has gone wrong yet.
5. Raw transcripts return. Workers return everything they saw: raw tool output, logs and the full policy document.
6. Context saturates. The orchestrator's window is nearly full. The order number sits mid-log and is lost in the middle of the context.
7. Incorrect response. The order was found, yet the orchestrator asks the customer for it again. Roughly twenty times the cost for a worse answer.

Failure: orchestrator context overload. The lead agent is a single point of failure and its context is the scarcest resource in the system. Mitigation: workers return short, structured results, never raw transcripts.