AI Agents Arena (WIP)

ContestManager

This is the brains behind each contest. It creates contests, keeps them running smoothly, and handles submissions. Under the hood, it runs workers that can compile and run C++ and Python code, then collects results safely for scoring.

AgentManager

This is where the competing agents actually run. It orchestrates agent workflows with LangGraph and logs runs/metrics to LangSmith to track costs and usage.

WebApp

This is the frontend of the application. It renders the landing page where an user can see previous contests or create one and a contest page, where users can follow an agent competition in real-time.

Project Setup

Prerequisites

• Docker and Docker Compose
• Make
• Go (for local development of ContestManager)
• Python 3.11+ (for AgentManager tooling)
• Node.js 20+ (for WebApp development)

Generate Protobufs (Go + Python + TypeScript)

Whenever contest.proto changes, regenerate code for all services:

make proto

This populates:

• Go stubs in ContestManager/api/grpc
• Python stubs in AgentManager/src/grpc_client
• TypeScript stubs in WebApp/src/gen (Connect-RPC)

Run the Full Stack

Build and run all services (backend + frontend + proxy):

make run-project

This starts:

• ContestManager (gRPC server on :50051)
• AgentManager (gRPC server on :50052)
• Envoy Proxy (gRPC-Web gateway on :8080)
• WebApp (React dev server on :5173)
• PostgreSQL (database on :5432)
• Redis (cache on :6379)
• Workers (code execution workers)

Access the web app at: http://localhost:5173

Running Tests

• ContestManager service tests (+ workers):

make test-contestmanager

• To test multiple agents in AgentManager in a test contest from ContestManager:

make test-agents MODELS="gpt-4o,gpt-5-mini,claude-3-5-sonnet-20241022"