The Risks of Multi-Agent Hiring Pipelines

As organizations increasingly adopt multi-agent systems for automated résumé screening, the interaction between these agents introduces new failure modes that go beyond simple single-model bias. The research highlights two critical phenomena: Access Bias and Recurrence Bias. Access bias refers to the initial filtering stage where agents may systematically exclude qualified candidates based on non-meritocratic features. Recurrence bias occurs when the output of one agent becomes the input for another, creating a feedback loop that reinforces initial errors and narrows the candidate pool over time.

Evaluating Agent-to-Agent Interaction

The study demonstrates that when agents operate in a sequence—such as a 'screener' agent passing candidates to a 'ranker' agent—the cumulative effect is not additive but multiplicative. If the first agent has a slight preference for specific linguistic patterns or demographic proxies, the second agent often amplifies this preference rather than correcting it. This results in a 'homogenization' of the candidate pool, where the system converges on a narrow set of profiles, effectively locking out diverse talent that might have been identified by a human recruiter or a single, well-calibrated model.

Implications for AI-Powered Recruitment

For builders of AI-powered hiring tools, the findings suggest that 'more agents' does not equate to 'more intelligence' or 'more fairness.' The authors argue that developers must implement explicit audit points between agents to prevent the propagation of bias. Relying on autonomous agents to 'reason' through résumés without human-in-the-loop validation or strict constraint-based filtering leads to systemic exclusion. The research emphasizes that the lack of transparency in how agents pass information to one another makes it difficult to debug why certain candidates are rejected, necessitating a shift toward more interpretable agent architectures in high-stakes domains like employment.