I use AI (Oct 2026)

A research engineer at Salesforce AI Research describes how their AI use has shifted over a year, from coding assistance toward delegating much of the work to autonomous agents. They now oversee five or six agents that implement features, reproduce bugs, run tests, and prepare pull requests, enabling fast prototypes, ambitious experiments, and less tedious work. But agents still produce bloated, poorly designed code without human judgment; the engineer’s most effective work included replacing thousands of generated lines with a compact tool they wrote themselves. Managing multiple agents is also mentally draining and can erode the pleasure and craft of programming. They review agent work for architecture and recurring patterns, while using automated review to catch some correctness issues - not as a substitute for human oversight. Current model capability is already sufficient for effective agent coding, they argue, and open-source alternatives make dependence on the most expensive systems less concerning.

The engineer is developing a personal AI policy because everyday norms for using these tools remain unsettled, and AI can either support or undermine creative work with others. AI summaries have become the default for ordinary web searches and are often useful, though they can return wrong or incomplete information. Product recommendations remain unreliable, especially for consequential purchases, and ads in chat interfaces feel intrusive. For casual fact-finding, AI is convenient, though the engineer sometimes prefers to reason things out socially. They are especially enthusiastic about using language models as interactive teachers: examples include learning model fine-tuning, rehearsing emergency responses, brainstorming voice-agent techniques, and unpacking legal documents.