How We Use AI Coding Agents (www.daft.ai)

Last week, a team shared insights from a session on optimizing workflows with AI coding agents like Claude Code and Cursor. A key takeaway was the importance of thorough research before coding, especially for complex features that affect multiple parts of a codebase. Teams are encouraged to use AI tools like Claude Code to plan and architecture discussions, thereby preparing a coherent roadmap before execution. This planning helps build the vocabulary needed to communicate effectively with coding agents and results in more efficient coding outcomes. The discussion also highlighted the contrasting strengths of different AI tools; while Claude Code excels in complex, agentic tasks, Cursor shines in speed for isolated changes. Engineers shared experiences where AI excels at simple bugs, boilerplate code, and tests, but struggles with more complex tasks like authentication and concurrent systems. To maintain control over the coding process amid the rapid changes AI generates, the team recommended structured workflows, such as using plan modes, incremental commits, and multiple active sessions for parallel tasks. Additionally, they emphasized the ongoing necessity for human oversight, particularly in architectural reviews, ensuring that AI supports but does not replace human judgment in software development.