BuildAgent

Building open-source software from source is a deceptively hard problem. Repositories ship with diverse build systems, undocumented dependencies, version-sensitive toolchains, and environment assumptions that rarely match a fresh machine. Developers and researchers routinely lose hours to dependency resolution, missing configuration, and trial-and-error compilation before any real work can begin.

We propose BuildAgent, an LLM agent for autonomous repository building. BuildAgent ingests an arbitrary GitHub repository, reasons about its structure, identifies the relevant build tooling, resolves dependencies, and drives the project to a successful build with no human in the loop. By leveraging the reasoning and context-understanding capabilities of modern LLMs, the agent adapts to projects spanning different languages, build systems, and ecosystems.

The result is a foundation for any downstream workflow that depends on a working build: reproducible research, automated testing, large-scale program analysis, and tooling that operates over a built artifact rather than just source code. BuildAgent removes the repetitive setup tax so developers and automated systems can focus on what the code actually does.