The tested Task API configuration with PubMed and ClinicalTrials.gov access reduced the reported error rate from 11.8% to 7.6% across 288 questions. The exact thesis of this note is: Specialized data connectors can improve research-agent accuracy by narrowing retrieval and exposing records in structured form, but this evaluation establishes a result only for Parallel’s tested biomedical tasks.
The observed facts are specific. Parallel ran the same agent and model twice per question, once with the public web and once with the web plus connectors; each setup cost $0.10 per question. In the connector condition, reported accuracy rose from 87.5% to 92.9% on LitQA3 and from 89.2% to 91.7% on TrialQA. The mechanism is that a domain search can resolve look-alike trial names or surface a hard-to-find paper, while a structured record can make fields easier to retrieve and count than a dense web page. Domain connectors can narrow the search space for ambiguous or hard-to-find sources and return structured records that make specific fields easier to retrieve and count; they can complement rather than replace web search.
Why now: Parallel has made these connectors available and is using a benchmark to argue for their practical effect, shifting the discussion from connector availability to measured task performance. The first implication is: research-agent teams should test domain connectors against their own task mix and verify whether accuracy gains justify the added retrieval path. The evidence boundary is: this is a company-published evaluation on two biomedical LABBench2 subtasks, with free and well-indexed databases; it does not establish comparable gains in law, finance, or other settings. The first watch condition is: look for independent replications and evaluations on licensed or less-indexed data, with error analysis and costs measured beyond the per-question API price.