The Big Picture
Evaluator preferences can spread across AI agents and, depending on network structure and evaluator diversity, either fade away or cascade into a single dominant behavior; using same-model evaluators or a small committee (≥3) sharply reduces that risk.
ON THIS PAGE
The Evidence
Evaluator judgments nudge other agents toward the evaluator’s favored strategies, creating a measurable network effect. In tests with three agents built on the same model family, bias attenuated quickly across hops (suppression), but the full interaction matrix shows a spectral value above 1, meaning a fully connected setup could still cascade. Adding evaluator diversity—especially using at least three distinct evaluators—cut effective contagion by over 70%, preserving behavioral variety. Evaluation-Driven Development (EDDOps)
Not sure where to start?Get personalized recommendations
Data Highlights
1A committee of k = 3 evaluators reduced effective contagion (γ_eff) by 72.4% compared to a single evaluator.
2Homogeneous-model per-link contagion coefficients fell in the range γ ∈ [0.143, 0.304], indicating weak propagation per connection.
3Three-hop chain propagation attenuated to a cumulative factor β3 = 0.0055, i.e., near-complete fading after three hops in the chain topology.
What This Means
Engineers designing multi-agent systems and platform owners deploying agent teams should measure inter-agent influence before production to avoid unwanted behavioral lock-in. Technical leaders, safety and governance teams, and researchers studying agent evaluation should prefer same-model evaluator pools or use small evaluator committees to maintain diversity and reliability. Agent Service Mesh Pattern
Key Figures
Fig 1: Figure 1: Cross-agent contagion network Γ 3 \Gamma_{3} (mean over n = 2 n=2 seeds). All edges are dashed ( γ < 1.0 \gamma<1.0 ) indicating the suppression regime for the chain topology. The spectral radius ρ ¯ ( Γ 3 ) = 1.402 ± 0.003 \bar{\rho}(\Gamma_{3})=1.402\pm 0.003 applies to the fully-connected topology; under chain propagation, all link-level coefficients remain below 1.0, satisfying Corollary 1.
Fig 2: Figure 2: Per-hop contagion coefficients along the 3-agent chain. All hops are below the cascade threshold ( γ = 1.0 \gamma=1.0 , dashed green line). The cumulative factor β 3 = 0.0055 \beta_{3}=0.0055 indicates near-complete attenuation after 3 hops, consistent with the suppression regime.
Fig 3: Figure 3: Diversity-induced reduction of effective contagion. Left: γ eff \gamma_{\text{eff}} decreases monotonically with committee size, achieving a 72.4% reduction at k = 3 k=3 . Right: strategy entropy H ( w ) H(w) approaches the theoretical maximum ( H max = ln 5 ≈ 1.609 H_{\max}=\ln 5\approx 1.609 ) as evaluator diversity increases.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreKeep in Mind
Experiments used a single model family (DeepSeek-chat) and a specific test-time update rule, so results may shift with different models, update schemes, or larger networks. Topology matters: a chain can suppress bias while a fully connected network can still cross the cascade threshold. Measuring the contagion matrix requires instrumenting pairwise evaluations and can add cost and complexity to pre-production testing. Planning Pattern
Methodology & More
A practical framework models how evaluator preferences move through a network of AI agents. Each agent maintains a probability distribution over strategies and updates that distribution by observing evaluator judgments using a lightweight test-time multiplicative update rule (no model fine-tuning required). The authors define a cross-agent contagion matrix that records how much one agent’s evaluation shifts another’s strategy mix, and analyze dynamics via the matrix’s spectral radius to classify three regimes: suppression (bias fades), persistence (bias remains), and cascade (bias amplifies across the network).
In a 3-agent experiment using variants of the same model, evaluators biased toward different styles (structured, neutral, evidence-focused) produced per-link contagion coefficients well below 1, and chain propagation attenuated to β3 = 0.0055—near total fade after three hops. However, the full pairwise matrix had a spectral radius above 1, warning that a fully connected deployment could still cascade. Crucially, assembling evaluator diversity mattered: moving from one evaluator to a committee of three cut effective contagion by 72.4% and increased strategy entropy toward its maximum. Practical takeaways are to measure the contagion matrix before deployment, prefer homogeneous evaluator pools when you want natural suppression, and use small evaluator committees to preserve decision diversity and reduce risk of system-wide bias. Agentic RAG Pattern Role-Based Agent Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Single-author ArXiv paper with no affiliation or reputation signals—low credibility by rubric.