On Programming with Agents (zed.dev)

Zed has recently adopted AI agents into its software development workflow, highlighting a shift in how programming is approached within the company. Over the past few months, the team has been learning to harness the capabilities of large language models (LLMs), leveraging them as tools for automation while emphasizing the importance of human oversight. The core principle behind effective agent usage revolves around constraining the output of LLMs through careful prompting, allowing developers to focus on the design and strategy of their projects rather than just the coding process. The significance of this approach lies in the acknowledgment that while LLMs can streamline coding tasks, they cannot replace human judgment in defining project goals or ensuring code quality. Zed's team has developed three key rules to guide their interactions with AI agents: first, plan thoroughly before asking for code generation; second, remain engaged during the agent's process to catch potential issues early; and third, maintain rigorous scrutiny of the agent’s output, treating it as an external contribution that must meet the team’s standards. By integrating these practices, Zed aims to elevate their code quality and maximize the benefits of AI, allowing their developers to dedicate more time to the craft of software engineering.