Two-Agent Resume Screening Changes Who Reaches Human Review

Summary

Hiring screening is often modeled as a single automated judgment over a resume-job pair, but this study evaluates a bilateral alternative in which employer-side and candidate-side agents exchange evidence and revise their views before deciding who advances. Using 600 constructed resume-job pairs, the authors compare the procedures with GPT-5.5 and Claude Opus 4.7. The two-agent process raises overall advancement from 33.3% to 39.3% with GPT-5.5 and from 34.0% to 35.5% with Opus 4.7. On a common borderline set of 191 pairs run three times, pass-instance rates increase from 4.5% to 26.2% and from 6.5% to 16.1%, respectively. The change is not simply a more permissive threshold: the two-agent procedure both rejects some applications that one-call screening advances and advances some that one-call screening rejects. At similar pass volumes, the procedures select different applicants, and no one-call threshold consistently recovers the applications selected by the two-agent method. In fresh reruns of discovery-selected cases, two-agent-only selections recur less often than selections shared by both procedures under GPT-5.5; the same pattern is less certain with Opus 4.7. A separate one-call follow-up does not show a comparable decline. The authors conclude that when both sides of hiring are represented by agents, the screening procedure itself influences who reaches human review and how reliably that access recurs.