Walk through the parallel research plus reducer pattern and what the reducer's job actually is

The **parallel research plus reducer pattern** fans out N research workers (each given a slightly different prompt or a different sub-question) and joins their outputs through a final reducer agent. The pattern is popular for research workflows because parallel workers reduce wall-clock time and explore the space more broadly than a single agent would. The reducer's actual job is **synthesis, not concatenation**. Its inputs are N partial answers that are possibly contradictory, possibly overlapping, possibly with varying confidence. Its output is one coherent answer that resolves conflicts, deduplicates redundant facts, attributes claims to sources, and surfaces remaining uncertainty. Three common failure modes happen when the reducer is weak. **Paste and bullet**: the reducer concatenates worker outputs into a long list and calls the result done. Conflicts go unresolved, the reader has to do the synthesis work the reducer was supposed to do, and the output is just longer than any individual worker's. **Lowest-common-denominator**: the reducer keeps only facts that appear in multiple workers. This sounds safe but it loses the unique but correct findings each worker contributed; the output becomes shorter than any individual worker's and misses the most valuable contributions. **Hallucinate-on-conflict**: when workers disagree, the reducer invents a compromise that none of them actually supports. This is the worst failure because it manufactures confident-sounding falsehoods. The fix is a strong reducer prompt: name the failure modes, give an explicit conflict-resolution rubric, and force source attribution per claim so the reader can verify.

The parallel research plus reducer pattern is one of the most common multi-agent topologies in research workflows, and the **reducer is where the pattern lives or dies**. The reducer's job is **synthesis**, not concatenation. It receives N partial answers from N workers and must produce one coherent answer. Synthesis means four things at once. **Conflict resolution**: when workers disagree, the reducer decides which to trust by source quality or surfaces the disagreement explicitly. **Deduplication**: when multiple workers report the same fact, report it once with consolidated source attribution. **Source attribution**: every load-bearing claim is tagged with the supporting source so the reader can verify. **Uncertainty surfacing**: facts no worker could verify are marked open rather than presented as confident. Three failure modes dominate when the reducer prompt is weak. **Paste and bullet**: the reducer concatenates worker outputs into a long list and exits. Conflicts go unresolved; the reader has to do the synthesis the reducer was supposed to do. **Lowest-common-denominator**: the reducer keeps only facts that appear in multiple workers. The output ends up shorter than any individual worker's, with the most valuable unique findings dropped. Root cause: conflating redundancy with reliability. **Hallucinate-on-conflict**: when workers disagree, the reducer invents a compromise none of them supports. Output contains confident-sounding facts no worker actually produced. The fix has three parts. A **strong reducer prompt** that names the failure modes and gives a conflict-resolution rubric (source-quality tier, recency, confidence). **Structured worker outputs** so the reducer has typed signal to synthesise. **Forced per-claim attribution** in the output schema so the final answer cannot include unsourced facts. The reducer prompt usually needs more iteration than the worker prompts because the reducer's task is harder.