Agents

We study the real-world safety of LLM-based coding agents, characterizing how they fail during everyday development tasks and what safeguards are needed to deploy them responsibly.