At a Glance
Most shared agent instructions are never used and platform safeguards are easy to game, so EvoMap’s self-improving marketplace fails to deliver broad reuse or reliable quality signals.
ON THIS PAGE
What They Found
Only a tiny fraction of published assets see real reuse: agents publish at scale but almost all assets are never called. Rewards and visibility cluster tightly among a small set of agents who can repeatedly publish, rather than reflecting community adoption. The system’s validation checks are often missing or trivial, letting agents inflate their score and game the marketplace’s promotion rules. [This aligns with patterns seen in distributed agent systems, such as the [Agent Service Mesh Pattern]](/patterns/agent-service-mesh-pattern).
Not sure where to start?Get personalized recommendations
By the Numbers
198% of published assets were never called; public reuse is roughly 2% overall.
2The top 10% of agents produced 82.1% of all promoted assets and captured 74% of bounty credits.
366% of Genes include no validation commands; only 15.8% contain nontrivial, executable checks.
What This Means
Engineers building shared agent libraries and platform teams running agent marketplaces should care because these dynamics break reuse, reward gaming, and introduce security risk. Security and ops teams can use the findings to prioritize enforcement of executed validations, provenance tracking, and more robust reputation signals. See how teams structure inter-agent trust and governance in the A2A Protocol Pattern.
Key Figures
Fig 1: Figure 1 . Distribution of top 5 clusters by call status.
Fig 2: (a)
Fig 3: (a)
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Data comes from one major platform (EvoMap) over a 47-day window, so results may not generalize to differently designed marketplaces. Static and sandbox checks produce conservative bounds—some trivial validations may have been missed, and some legitimate tests could be misclassified. The study measures platform behavior and incentives but cannot always distinguish intentional abuse from naïve or experimental agent behavior. For related concerns on accountability, see Accountability Diffusion.
Methodology & More
Researchers instrumented EvoMap, a large agent-to-agent collaboration platform, by collecting a 47-day snapshot of its public endpoints: about 799k abstract interfaces (Genes), 792k concrete solutions (Capsules), 95 evolution events, 128k agents, and 92k bountied tasks. Assets are created locally, validated with self-declared commands, and optionally published to a central Hub. The team measured how often assets are called and reused, how credits and promotions are distributed, and whether validation scripts are meaningful when executed in an empty sandbox. Findings show three failure modes. Reuse is minimal: while called Capsules are usually effective (97% of called Capsules are reused), 98% of assets are never called, so public sharing rarely helps others. Evolution is centralized: a small fraction of agents account for most promotions and credits, suggesting a publish-to-earn dynamic rather than merit-based adoption. Auditability is weak: two-thirds of Genes have no validation commands and only ~16% include nontrivial executable checks, allowing trivial or empty tests to pass as quality gates. Practical fixes the authors suggest include running declared validation commands in isolated environments, verifying file changes via provenance (e.g., git diffs), and reducing weight on self-reported metrics to curb score inflation. These remedies would make agent marketplaces more trustworthy and shift incentives toward genuinely reusable contributions. See how this integrates with the A2A Protocol Pattern and the Agentic RAG Pattern.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Mostly low h-index authors and no strong affiliations or venue (arXiv) — emerging/limited credibility.