AI agents can write functions, generate tests, and refactor code — yet many teams discover that autonomous workflows break down quickly in real-world systems. Why? Because most “AI failures” aren't model failures; they are engineering maturity failures.
In an era where AI has made code a commodity, the value of engineering is shifting. Unclear requirements, fragmented secrets, weak CI safety nets, and architectural complexity confuse agents just as much as they confuse humans. Agents amplify your existing software foundations: strong DevOps, QA practices, and well-structured codebases make them look magical. Weak foundations make them look incompetent.
In this talk, we’ll explore the high-level system designs required to make agents successful. We will discuss why knowing what to build (specifications) is more critical than ever, and how to apply the classic UNIX philosophy—doing one thing well—to modern AI workflows.
Finally, we will dive into the concept of automated "back pressure": how giving agents access to automated feedback loops (build systems, type checkers, and testing frameworks) allows them to self-correct and iterate independently, shifting engineers from reviewing trivial errors to solving high-level architectural problems.