Your team is building a multi-agent research system for a finance class project. A coordinator agent dispatches to three subagents: a web searcher, an SEC filings analyzer, and a synthesis agent that writes the final report. After running it on the prompt "Analyze Tesla's competitive position in 2025," the report is coherent but the synthesis agent has invented two statistics that don't appear in any of the source materials.
Your teammate proposes adding a "fact-checker" subagent that reviews the synthesis output and flags unsupported claims.
Is this the right fix?
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.
The hallucinations aren't happening because there's no fact-checker. They're happening because the synthesis agent is receiving compressed prose from upstream agents, has lost the link between claims and their sources, and is filling gaps in its understanding by generating plausible-sounding statistics. The structural fix is upstream: require the web searcher and filings analyzer to output structured claim-source mappings (claim, source URL or document, supporting excerpt, date), and require the synthesis agent to preserve those mappings. When the architecture preserves provenance, the model has nothing to hallucinate around.
Adding a fact-checker subagent is tempting because it sounds like defense-in-depth. But it adds latency, cost, and a new failure mode (the fact-checker can also be wrong), without addressing the upstream issue. You'd be papering over a structural problem with another agent. Often the right fix in multi-agent systems is better information flow, not more agents.
Model selection rarely fixes architectural information loss. A stronger fact-checker still can't verify what the synthesis agent has invented if the source material it would check against is no longer in the system.
This is the defeatist answer. Hallucination in multi-source synthesis is largely an architectural problem, not a model limitation. Systems that preserve provenance and require source attribution dramatically reduce hallucination. Disclaimers don't help users trust the output more — fixing the architecture does.