We’ve all asked an AI agent to figure out how an unfamiliar codebase works, only to wonder whether it actually found the answer or hallucinated it.

What if the agent could show you exactly where it found the answer? It could explore the codebase, find the relevant implementation, and return an answer backed by citations to the exact file names and line numbers.

In this talk, we’ll discuss how we built Megasthenes, a code research agent that takes a plain-English question about a codebase and returns evidence-backed answers. We’ll look at how we built its sandbox, assembled the tools it needs to research code, and used evals to understand and improve its performance.

By the end of the talk, you’ll understand the key design decisions behind building a code research agent and the sandbox it operates in.