Code Coliseum

Code Coliseum addresses a new problem created by the rise of AI coding agents: generating code is becoming easier, but deciding which AI-generated implementation should actually be trusted and merged is still largely manual. Developers can ask multiple AI agents to solve the same task, but their answers may use different architectures, trade-offs, levels of complexity, and runtime behavior. A convincing explanation does not necessarily mean the implementation is correct, safe, or maintainable. Code Coliseum turns this problem into a controlled engineering competition. A developer submits one feature request. Three IBM Bob coding agents — Minimalist, Sprinter, and Architect — receive the same task but approach it with different engineering strategies. Each agent works independently inside an isolated Git worktree created from the same base commit, preventing contenders from modifying the developer’s main working tree or influencing one another. Once implementation is complete, every contender enters the same deterministic Gauntlet. Code Coliseum validates dependency integrity, TypeScript safety, production build success, and required runtime behavior. A contender that fails any mandatory check is eliminated. Survivors are then compared using real engineering evidence including files changed, lines changed, execution time, Gauntlet results, and the exact Git diff produced by each agent. The developer can inspect those changes, select the implementation that best fits the project, and apply that exact candidate back to the working tree. Code Coliseum intentionally keeps the developer in control. It does not automatically commit or push generated code. The project is designed for developers and engineering teams adopting agentic software development who need a more reliable way to evaluate AI-generated implementations. What makes Code Coliseum different is that it does not try to be another coding agent. It is an evaluation arena for coding agents. Instead of “generate and hope,” Code Coliseum creates a workflow of compete, prove, inspect, and merge. Make your AI earn the merge.