A supervisor/subagent case study — an orchestrator that decomposes a research topic, parallel researcher subagents, a writer agent that synthesizes one coherent cited report, a worked cost/latency calculation, and failure handling for a subagent that comes up empty.
The project brief: take a research topic broad enough that no single search answers it, and produce a structured, cited report covering it comprehensively — fast. A single agent researching sub-topics one at a time in sequence is slow and tends to lose coherence over a long session, so this build splits the job three ways: an orchestrator agent decomposes the topic into well-scoped sub-questions, researcher subagents investigate each one independently and in parallel, and a writer agent synthesizes their findings into one report with a consistent voice and no duplicated content.
This module walks the coordination design a production AI engineer actually has to get right: why a supervisor topology fits this task better than peer-to-peer negotiation, how the orchestrator's own prompt is engineered to avoid both overlapping and gap-leaving sub-questions, what gets passed between researcher subagents and the writer agent (and why raw transcripts instead of clean summaries commonly blow the context budget), a worked cost-and-latency calculation comparing parallel and sequential dispatch, and how the system handles a subagent whose search comes up empty. The full worked numbers, complete tech-stack manifest, and finished design reasoning are part of the unlocked module below.