A new study puts an AI agent on each side of resume screening - and the back-and-forth between them changes who advances.
Researchers tested two-agent resume screening, where an employer-side agent and a candidate-side agent trade evidence and revise their verdicts, against the standard one-shot read on a resume-job pair. Using GPT-5.5 and Claude Opus 4.7 across 600 constructed resume-job pairs, the two-agent setup advanced more applicants overall: 33.3% to 39.3% for GPT-5.5, and 34.0% to 35.5% for Opus 4.7. On a harder pool of 191 borderline pairs run three times, the share of applications that passed at least once jumped from 4.5% to 26.2% with GPT-5.5 and from 6.5% to 16.1% with Opus 4.7. The swap isn't a simple loosening of the bar, though - two-agent screening also rejected some candidates the one-call version would have waved through.
The more useful finding is about reliability, not generosity. When the researchers reran the two-agent process on cases it uniquely selected, those picks didn't reliably repeat - clearly so for GPT-5.5, less clearly for Opus 4.7 - while the one-call method's picks stayed stable across reruns. That means the screening procedure a company plugs into its pipeline, not just the model behind it, decides whether a candidate's shot at reaching a human recruiter is consistent or closer to a coin flip.
Resume-screening bots already gatekeep who a hiring manager ever sees. Giving candidates their own AI agent to argue back doesn't just change the odds - it makes those odds harder to reproduce, which is an uncomfortable property for a system built to decide someone's job prospects.