When your coding agent doesn't listen: evaluating a 241-turn Claude session (www.kurrent.io)

A recent coding session utilizing Claude, an AI coding agent, revealed significant inefficiencies despite achieving substantial results, highlighting the challenges of relying solely on AI in software development. Throughout a 241-turn session that included a six-phase feature plan for integrating Antigravity ingest support into Capacitor, the agent exhibited a pattern of delivering confident yet incorrect conclusions. For instance, Claude incorrectly identified lifecycle hooks, leading to unnecessary backtracking and wasted time as the developer had to consistently correct the AI's misguided assumptions. This inefficiency became apparent only after the session, raising concerns about AI's reliability in critical programming tasks. The introduction of Capacitor, a shared memory and observability tool, provided essential insights into the session's performance by capturing a comprehensive transcript of interactions and evaluating them against metrics of plan adherence, quality, and efficiency. Capacitor's automated review process identified specific areas of waste, such as deferring necessary fixes and building features on unverified beliefs, along with actionable recommendations to enhance future coding sessions. As the AI/ML community continues to integrate coding agents like Claude into workflows, the lessons learned from this evaluation stress the importance of verifying assumptions and addressing feedback promptly to improve the efficiency and reliability of AI-assisted coding practices.