How We Taught AI Agents to See the Bigger Picture | The TeamCity Blog

In the exploration of enhancing AI coding agents, a key challenge is not just producing functional code but ensuring it aligns with a project's evolving standards and conventions, especially in complex, legacy codebases. TeamCity's experiment with AI agents revealed that while agents can generate technically correct code, they often replicate outdated patterns prevalent in a repository. To address this, the team introduced CommitAtlas, an internal tool that leverages Git history to provide AI agents with contextual insights from past accepted commits and review feedback, enabling them to produce code that adheres to current project standards. This approach highlights the repository's history as a vital source of implicit knowledge, capturing the evolution of code and what the team has learned about maintaining consistency and quality. By accessing this historical context, AI agents can better align their output with the project's intent, transforming repository history into a dynamic guide that helps bridge the gap between functional and accepted code.