Field notes from AI-native engineering
Short reads from the Meta.Dev team. New posts most days.
Traditional assert-equals testing falls apart when the output is a paragraph the model wrote. Here's the pattern that actually scales.
Model Context Protocol gets buried under demos. Here's the EM-level picture of what it changes about how your team ships.
We stopped running leetcode rounds two years ago. The loop that actually catches AI-native depth looks like this.
With million-token context windows on the table, half of yesterday's RAG pipelines were premature. Here's how we decide.
AI-native engineers don't write more code. They run more parallel work. Here's what that actually looks like.
Not every "AI workflow" is an agent. Mapping the rungs from autocomplete to autonomous and where most teams should actually live.
Most teams hire engineers and then bolt AI onto their workflow. AI-native teams flip that order — and the gap is starting to show.
The job of a senior engineer is changing. The first draft of the code is no longer where the value is — it's the verification, the decomposition, and the architecture.
Most teams reach for fine-tuning when they should reach for retrieval. A short decision rule we use on every engagement.