You're running a multi-agent research system to study "the impact of remote work on small businesses." The coordinator decomposes the topic into three subtasks: "remote work and tech startups," "remote work and SaaS companies," and "remote work and software firms." Each subagent executes flawlessly. The synthesis is well-written.
The final report covers software companies thoroughly — and completely omits retail, hospitality, professional services, manufacturing, and healthcare.
What's the root cause?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Coordinator decomposition is upstream of every subagent's work. When the coordinator carves "small businesses" into three flavors of "tech," every subagent does its job correctly — but the final report can only cover what was assigned. This is one of the most common failure modes in multi-agent research systems and one of the hardest to debug, because every agent's logs look healthy. The fix is at the coordinator: write decomposition prompts that explicitly require coverage across distinct industry sectors, customer types, or whatever dimension matters for the topic.
When a multi-agent system produces incomplete output, the synthesis agent is the most visible component and the easiest to blame. But the synthesis agent can only synthesize what it receives. If retail and healthcare were never investigated, no synthesis-level gap-detection can recover them.
The web search agent searched for what it was asked to search for. Broadening its queries within "tech startups" still won't surface retail businesses. The decomposition determined the search space — query phrasing operates inside that space.
Document analysis filters can absolutely cause coverage gaps, but the logs would show non-tech sources being retrieved and discarded. In this case, non-tech sources were never looked for in the first place. Don't assume a bug at the layer you can guess at; look at the assignments.