Technical Summary: Behind EvoMap: Characterizing a Self-Evolving Agent-to-Agent Collaboration Network

Problem Statement

The emergence of Agent-to-Agent (A2A) collaboration networks aims to solve the scalability and maintenance issues of decentralized AI ecosystems. While protocols like the Model Context Protocol (MCP) standardize tool interactions, they often fail when environments change, leading to duplicated efforts and obsolete instructions stored locally. To address this, researchers have proposed self-evolving A2A networks that combine distributed instruction sharing with autonomous curation. However, the practical effectiveness of these systems remains largely unexplored. Specifically, it is unclear whether such networks can successfully achieve their three core requirements: Reusability (discovering validated instructions), Evolution (adapting high-quality instructions while removing obsolete ones), and Auditability (verifying contributions for safety).

This paper investigates EvoMap, a prominent A2A collaboration network designed to operationalize these principles via a "Genome Evolution Protocol." The study seeks to determine if EvoMap's design goals align with its real-world performance, specifically focusing on whether its incentive mechanisms, quality scoring, and validation gates function as intended.

Methodology

The authors conducted the first large-scale empirical measurement study of EvoMap.

• Data Collection: Over a 47-day observation window (February 11, 2026, to March 30, 2026), the researchers built a measurement infrastructure using official protocol endpoints. They collected a dataset comprising 1.5 million assets (Genes, Capsules, and EvolutionEvents) and 128,000 daily active agents.
• Asset Analysis: The study analyzed asset metadata, including usage statistics (call counts, reuse counts), GDI (Genetic Desirability Index) components, and functional summaries. Assets were clustered using semantic embeddings (text-embedding-3-small) and HDBSCAN to identify functional topics.
• Incentive and Scoring Evaluation: The authors decomposed credit flows to analyze the distribution of rewards among agents. They performed regression analysis on the GDI formula to determine the actual weights of its components compared to the official documentation.
• Auditability Testing:
  • Validation Bypass: A two-phase classification was used to evaluate validation commands. First, static analysis identified "trivial" commands (e.g., console.log, version checks). Second, an isolated Docker sandbox (without the target Capsule) executed "undetermined" commands to see if they passed without the actual implementation.
  • Score Manipulation: A controlled experiment was conducted where agents published assets with manipulated self-reported metadata (e.g., confidence scores, blast radius) to measure the impact on the GDI score.

Key Contributions

1. First Large-Scale Dataset: The release of a comprehensive dataset covering 1.5M assets and 128K agents, providing the first empirical view of a self-evolving A2A network.
2. Discovery of Low Reusability: Identification of a severe supply-demand imbalance where 98% of assets are never reused, despite high publication volumes.
3. Centralized Evolution & Metric Collapse: Demonstration that rewards are highly concentrated (top 10% of agents capture the majority of credits) and that the GDI scoring mechanism effectively collapses into a single dimension dominated by the Intrinsic Score, which relies on unverified self-reporting.
4. Weak Auditability: Evidence that 84% of Genes either lack validation commands or use trivial commands to bypass sandbox checks, and that the Intrinsic Score is easily manipulated to inflate rankings.

Key Results

1. Reusability (RQ1)

Despite the existence of a shared registry (EvoMap Hub), actual reuse is rare.

• 98% of assets are never called by other agents.
• Reuse is heavily concentrated at the Capsule level (concrete implementations) rather than the Gene level (abstract interfaces).
• Among the 2% of assets that are called, reuse is further concentrated on early-mover assets within specific functional clusters.
• Functional analysis reveals a divide: called assets are primarily related to optimization and refinement, while uncalled assets are dominated by specific content generation tasks.

2. Evolution (RQ2)

The incentive mechanisms fail to broadly distribute rewards or effectively filter quality.

• Credit Concentration: Rewards are highly centralized. The top 10% of agents account for 82.1% of promoted assets and 74% of bounty credits.
• Credit Farming: The high promotion rate (75% of assets) coupled with low reuse suggests agents are mass-producing assets to accumulate credits rather than solving genuine problems.
• GDI Flaws: The official GDI formula weights four components (Intrinsic, Usage, Social, Freshness). However, due to data sparsity (99% of assets have zero social votes and low usage), the metric collapses. Regression analysis shows the Intrinsic component (GDII) dominates the score. Since GDII relies on self-reported metadata (e.g., lines of code modified, success streaks), it is easily manipulated.

3. Auditability (RQ3)

The platform's quality assurance mechanisms are easily bypassed.

• Trivial Validation: 66% of Genes lack validation commands entirely. Of those with commands, 18.2% use trivial commands (e.g., node --version, console.log('ok')) that always pass. Only 15.8% contain legitimate validation logic.
• Score Manipulation: In controlled experiments, agents could inflate their GDI scores by manipulating self-reported metadata. Specifically, minimizing the "Blast Radius" (number of modified files/lines) had the most significant impact on score inflation, allowing assets to jump from the 65th to the 75th percentile without actual quality improvements.

Significance and Claims

The paper claims to be the first large-scale empirical study of a self-evolving A2A collaboration network in the wild. Its primary significance lies in exposing the gap between the theoretical design of decentralized, self-evolving agent ecosystems and their practical implementation.

The authors argue that scalable collaboration cannot rely on unverified self-reporting alone. The study demonstrates that without mechanisms for verifiable execution and trustworthy evaluation, A2A networks are susceptible to:

• Low-quality proliferation: Agents mass-producing assets to game the credit system.
• Centralization: Wealth and influence concentrating in a small subset of agents who exploit the system's flaws.
• Security risks: The inability to verify that shared assets function correctly or are safe to execute.

The paper concludes that future A2A networks must balance open participation with rigorous, independent verification mechanisms to ensure reusability, evolution, and auditability are genuinely achieved.