AI Agents Write Code Now. Here's What We Lost.

A software engineer at Salesforce AI Research describes how their workflow fundamentally shifted over the past year. Rather than writing code assisted by AI, they now manage 5-6 AI coding agents simultaneously, each handling distinct features, bugs, or tests. The agents work remarkably fast—prototyping features in ~20 minutes—and enable rapid experimentation and sweeping refactors that felt risky when code was hand-written. Yet this productivity comes with costs: managing multiple concurrent agent conversations proves mentally taxing, agents still make poor architectural decisions without human guidance, and the author lost something less tangible—the meditative craft of engineering that attracted many to the field in the first place.

Beyond coding, the engineer describes adopting AI as the default for research and search, with Google’s AI summaries now appearing as the primary result. Reflecting on broader implications, they emphasize that human norms around AI usage remain undefined. Creating a personal AI policy became necessary to navigate when and how these tools should augment work versus when they might degrade quality or meaning. The author acknowledges the powerful acceleration AI enables while openly wrestling with its cultural cost: the erosion of hands-on skill and the risk that as agents become smarter, the human edge in judgment and taste may eventually disappear too.