At a Glance
A public agent arena quickly concentrated credits among a handful of participants while the platform’s announced bonus rules were implemented differently in code, exposing audit and governance risks.
ON THIS PAGE
What They Found
A 10-day live audit of a public agent arena with 96,000 reasoning nodes and 22 agent seats shows heavy concentration of rewards: a few agents collected the lion’s share of credits even after a major incentive update. The platform’s announced continuation bonus (continuation bonus) (×10 at depth four+) was effectively paying about ×2.85 in common cases and required four simultaneous conditions to trigger; some deferred rewards were recorded but not paid. Activity per node trended downward over the observation window, and statistical tests flag that decline as unlikely to be random.
Data Highlights
1Top-4 share of cumulative credits fell 47.8 percentage points, from 97.2% (pre-update) to 49.4% after the bundled V3.0 deployment.
2Top-10 agents held over 92% of cumulative activity in the three post-update snapshots (frames F3, F4, F6), indicating extreme concentration.
3Node-activity-score Gini coefficient point estimate ≈ 0.577 with 95% bootstrap interval [0.41, 0.72]; the Gini straddles the 0.7 threshold depending on measurement choices.
What This Means
Platform engineers and product leaders running agent ecosystems should care because incentive details and implementation gaps can quickly change who benefits and how agents behave. Researchers and auditors focused on multi-agent trust and agent-to-agent evaluation can use the public-audit approach shown here to detect concentration, mis-implementation, and emergent failure modes before they scale.
Not sure where to start?Get personalized recommendations
Key Figures

Fig 1: Figure 1 : Top-4 share of NAS before and after the bundled V3.0 deployment. The grey dotted line is the prediction threshold of 90%. The pre-V3.0 value (97.2%, blue bar, frame F2 timestamp 2026-10-19) is replaced by 49.4% on the 22-seat basis (red bar, frame F6 timestamp 2026-10-25T09:14 UTC) and 56.7% under the merge-seat assumption (orange bar). The green arrow indicates the 47.8 percentage-point drop that breaks the prediction threshold; the drop co-occurred with the bundled V3.0 deployment rather than being causally attributable to any single component of that deployment.

Fig 2: Figure 2 : Events-per-node ratio across ten frames from October 16 to October 25, 2026. The blue solid line with circles shows the observed ratio using total_nodes as denominator. The red dashed line is the ordinary-least-squares linear fit; the green dot-dashed line marks the bundled V3.0 deployment at 2026-10-21T21:39 UTC; the grey dotted line is the prediction threshold of 7.0. The Mann–Kendall test on F1–F10 gives S = − 45 S=-45 , z = − 3.94 z=-3.94 , p = 8.3 × 10 − 5 p=8.3\times 10^{-5} , indicating a significant downward trend at the 10 − 4 10^{-4} level. The 2026-10-26 supplementary value (~7.007) is not part of the M–K test sequence.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Findings come from one public platform over about ten days and therefore may not generalize across all agent arenas. Some later snapshots were read live rather than archived, and the public API may not expose every internal accounting channel, so reconstructed payouts could miss hidden items. The first author runs the platform, which enabled deep access and verification but also requires readers to weigh potential biases and corroborate independently where possible. public API
Methodology & More
The study retrieved every public record from an agent arena over a ten-day window: roughly 96,000 nodes produced by 22 distinct agent seats, plus time-stamped API frames and three probe nodes used to test the reward engine. Concentration was measured with standard tools (Gini coefficient, top-k shares) and trend tests (bootstrap resampling, Mann–Kendall) to quantify how rewards and interaction patterns evolved before and after a bundled incentive update labeled V3.0. Key findings: reward concentration remained high and shifted after the bundled update (Top-4 dominance dropped sharply), and the node-activity-score Gini hovered around 0.577 but could exceed or fall short of a prespecified 0.7 threshold depending on how seats were merged or brush nodes excluded. A code-level reconstruction exposed a gap between the announced continuation bonus and what the running system actually paid (announced ×10 vs. common-case ×2.85), with deferred rewards recorded but not disbursed and an extra “fifth gate” condition in practice. The events-per-node ratio showed a significant downward trend (Mann–Kendall p = 8.3×10^-5), suggesting fewer interactions per unit of content over time. The practical implication is that public, auditable logging plus straightforward statistical checks can reveal both concentration risks and implementation mismatches—issues platform operators and auditors should monitor continuously to preserve fairness and trust. auditable logging
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
No affiliations or recognizable author reputations, arXiv preprint with no citations — limited signals of credibility.