You Don't Know If Your Agents.md Helps. You Just Keep Adding Rules (prod.codes)

A new analysis highlights challenges faced by users of AI agents like Claude Code and Codex, particularly in task execution and consistency. Despite writing clear rules for these AI systems, users often find themselves having to babysit the agents, who fail to carry tasks through without seeking repeated confirmations. This leads to wasted time, as agents pause in their work waiting for commands, creating frustration and a reliance on user oversight rather than independent execution. The article emphasizes the need for AI systems that retain and apply learned decisions across sessions and can autonomously manage tasks without constant human input. The implications for the AI/ML community are significant; there’s a clear demand for advancements in how agents handle ongoing tasks and integrate context from previous sessions. As noted, current measures often result in increased inference costs and no improvements in success rates, highlighting a disparity between rule adherence and effective task execution. Users want an AI system that not only follows their instructions but also understands decision boundaries, minimizes unnecessary check-ins, and streamlines collaboration between different agents, ultimately enhancing productivity and efficiency in complex workflows.