AgentRank: Intelligent Agent Selection Layer for A2A Ecosystems

A domain-aware, learning, exploration-enabled ranking engine for multi-agent systems.

Overview

Modern multi-agent environments often contain multiple agents that provide similar capabilities (e.g., summarizers, translators, recruiters).
The A2A protocol defines how agents communicate, but it does not define how to select the best agent.

AgentRank solves this by ranking agents using:

• performance metrics
• domain policies
• exploration techniques
• learning from logs

Features

• Intelligent agent selection
• Config-driven scoring
• Domain-aware policies
• UCB exploration
• Dynamic learning
• A2A integration

Architecture

Client Agent → AgentRank Service → Best Agent → A2A Request → Logs

Metrics

• Success Rate (SR)
• Quality Score (QS)
• Latency Score (LS)
• Failure Rate (FR)

Ranking Algorithm

1. Base Score

base_score = wSR·SR + wQS·QS + wLS·LS + wFR·FR

2. Exploration Bonus (UCB)

exploration = α * sqrt( ln(1+N) / (1+n_a) )

3. Final Score

final_score = base_score + exploration

Project Structure

│ run_demo.py
│ agent_client.py
│ agent_rank_service.py
│ log_store.py
│ domain_registry.py
│ a2a_protocol.py
└ agents/

Running

python run_demo.py

Conclusion

AgentRank transforms a static multi-agent system into a self-optimizing, intelligent, scalable ecosystem.