Agent Data Pipelines: Nightly S3 Snapshots Plus LLM ETL Instead of Scripts

curious_vii · x · 2026-08-24

The author proposes a new approach to consolidating team-wide agent data such as traces: take full machine snapshots every night into timestamped S3 storage, then run ETL via a high-power agent thread (Fable, Astra, etc.) rather than traditional scripts.

The core argument: the traces system of record can still be a relational DB, but S3 snapshots are needed for comprehensive coverage — with almost no durable standards in the space, agents will do weird dynamic stuff that breaks conventional software logic. This also raises interesting questions for BYOD contexts.

Original post →

More from coding & agent

coding & agent channel →