Some early & evolving thoughts on "agentic DX": 1️⃣ More 'hyperlocal reasoning' APIs (like Tailwind) i.e.: APIs that rely less on 'reading in more context' in order to be interpreted. 2️⃣ Less pressure on verbosity. While humans' fingers get tired when typing, LLMs emitting tokens do not. APIs like `𝚍𝚊𝚗𝚐𝚎𝚛𝚘𝚞𝚜𝚕𝚢𝚂𝚎𝚝𝙸𝚗𝚗𝚎𝚛𝙷𝚃𝙼𝙻` are good, actually. Explicit, convey semantics, etc. This is also true for logging verbosity, stack traces, etc. 3️⃣ Double down on the filesystem Claude Code-like agents are reinforced on filesystem operations. They like to `ls`, `find`, `cat`, `head`. Filesystem route organization in @nextjs App Router naturally 'sections off' parts of the application that can get their own specific agent instructions. 4️⃣ CLI ergonomics CLIs generally need to provide better feedback, have non-interactive modes, and centralize output. An example here is @nextjs forwarding browser logs (currently under a flag) so that agents can have a 'single pane of glass' of debugging info: Overall, there's a lot of overlap between "good DX" and "good agent DX". I think of making things better for agents as an optimization. If the foundations are not solid, no amount of AI will fix them.
question for @rauchg: how does AI change how you think about framework/API design from a “developer experience” PoV? do you now do things that are different assuming AI will implement? or does it reinforce everything you’ve regarded important since AI makes it more relevant?