Skip to content

Latest commit

 

History

6 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

MAS-GraphJudge

Measure how, not just whether

A graph-based coordination benchmark where a "Green Agent Judge" dissects multi-agent interactions into network graphs and evaluates collaboration quality through structural analysis and LLM assessment.

Version License CodeQL CodeFactor ruff pytest

Product Vision

GraphJudge transforms multi-agent interactions into coordination graphs and evaluates collaboration quality through graph theory, LLM assessment, and performance analysis.

Novel Value

Graph-based runtime coordination analysis - measuring collaboration quality through NetworkX metrics and LLM-as-judge evaluation.

Evaluation Tiers

Tier Type Description
1 Graph NetworkX metrics (centrality, efficiency)
2 LLM-Judge Coordination quality assessment
3 Text Similarity metrics (plugin for PeerRead)

Submission

Competition Tracks

  • Research Agent
  • Multi-Agent
  • AAA (Agentified Agent Assessment)

Abstract

See docs/GreenAgent-UserStory.md for full problem statement and value proposition.

TL;DR: Graph-based structural analysis + LLM-as-judge + text similarity for multi-tier coordination evaluation. A2A-compliant trace capture enables reproducible assessment (0% variance) of multi-agent collaboration patterns.

Demo Video

Status: Coming Soon

Demo video (max 3 minutes) will cover: server startup, evaluation flow, and results interpretation.

Roadmap

  • ✅ Phase 1: A2A + Graph + Basic eval (current)
  • 🔜 Phase 2 (outlook): ART training on traces, potentially using WeightWatcher or PerforatedAI
  • 🔮 Phase 3 (outlook): Self-evolving GreenAgent, e.g., DGM

Note: Time constraints limited full implementation of advanced features planned in Phase 2 and 3. Current release focuses on core graph-based evaluation with proven reproducibility.

Quick Start

Run from GHCR (no build needed)

docker pull ghcr.io/qte77/agentbeats-greenagent:latest
docker run -p 9009:9009 ghcr.io/qte77/agentbeats-greenagent:latest
curl localhost:9009/.well-known/agent.json

Developer Setup

See docs/AgentBeats/QUICKSTART.md for local development, Docker build, and environment configuration.

Competition

Document Purpose
COMPETITION-ALIGNMENT.md Alignment analysis with competition requirements
SUBMISSION-GUIDE.md Phase 1 submission checklist
LIMITATIONS.md Scope boundaries and deferred features
RESOURCES.md External links and references

Releases

Packages

Used by

Contributors

Languages