OUR FIRST OUTAGE FROM LLM-WRITTEN CODE

Summary from the TLDR newsletter

Sketch experienced a series of mini-outages on July 15. An LLM had made a small but critical change during a refactor that was missed by a human reviewer. When humans refactor code, they select the original text, cut it, move it to the new file, paste it, and then make intentional changes. LLMs write two patches, a deletion and an insertion, leaving room for transcription errors. Sketch added clipboard support to its agent environment to try to prevent further issues in transcription.