asherdale1

fair on length. it's more a 'how we built it' post, so most of it is mechanics (sandboxing grep and sed, catching conflicting writes) rather than data. the empirical bits are linked inline: the study on grep vs vector retrieval, and anthropic's docs on self-hosted sync silently replacing conflicting edits

no benchmarks of our own yet, but that's coming soon

feels like everyone is converging on durable agents lately: pi durable, cloudflare agents on durable objects, vercel's durable harness