A staff engineer shares how their LLM usage has evolved over the past 15 months. The biggest shift: AI agents are now good enough to draft entire PRs, investigate most bugs autonomously, handle testing, and assist with local setup tasks. The author now starts every code change by asking an agent to solve it first, then does a single editing pass. They still write all public communications (PR descriptions, ADRs, Slack messages) by hand and carefully review all agent-produced code. The key insight is that the core skill is finding the right balance — delegating as much as possible to agents without over-trusting them on things requiring human judgment or communication.
Sort: