A pipeline of three agents (Researcher → Analyst → Editor) is producing final outputs that subtly disagree with the original sources. The Researcher cites "X grew 12.3% YoY in Q3 2024." The Analyst writes "X showed strong growth recently." The Editor writes "X has experienced impressive expansion in the past year." By the time the user reads it, the precise statistic has been replaced with vague prose three times over.
What's happening, and how do you fix it?
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 "telephone game" problem: each handoff compresses, and compressions stack. Each agent is doing its job — paraphrase, summarize, polish — but the cumulative effect is precision loss. The fix isn't at the agent level; it's at the handoff level. Structured handoff formats preserve the literal claims and statistics through the pipeline. The Editor receives both the Analyst's synthesis and the original Researcher's claim-source mappings, and is evaluated on whether the final output preserves the precise facts. Provenance has to be preserved through structure, not trusted to free prose.
The Editor isn't inventing facts — it's receiving "strong growth recently" from the Analyst and producing prose consistent with that. The data wasn't there for it to preserve. Pinning it on hallucination misses where the precision actually got lost.
Treating systematic precision loss as "normal variation" is how reports end up subtly wrong in ways that erode trust. Compounding distortion is a structural property of unconstrained handoffs; it's exactly the kind of problem multi-agent design should solve, not accept.
Removing the Analyst skips the synthesis step entirely — now the Editor has to do both jobs (analyze + edit) which produces its own quality issues. The pipeline isn't broken; the handoff format is. Fix the handoff, not the headcount.