You're synthesizing a research report combining findings from multiple subagents. One source (a 2022 industry report) states the market grew 12% last year. Another source (a 2024 academic paper) states the market grew 8% last year.
What should the synthesis subagent do?
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.
When credible sources disagree, the synthesis output should preserve the disagreement, attribute each value to its source, and surface methodological context where possible. This respects the user's intelligence, makes the report defensible, and lets readers evaluate the sources themselves. Annotating conflicts is dramatically more honest than collapsing them, and it's the standard for any synthesis that aspires to credibility.
Recency is one signal but not a sufficient one. The 2022 report might use a broader market definition; the 2024 paper might use a narrower one. Picking based on a single dimension hides the underlying methodological question.
Averaging two estimates from different methodologies produces a number that doesn't correspond to either methodology's definition of the market. It looks like a compromise but is actually meaningless.
Dropping conflicting findings is intellectual cowardice and produces a report that's less informative than the underlying sources. Synthesis should compress; it shouldn't censor.