OpenAI's Agent-Swarm Playbook for Boring Spark Upgrades, Shown at QCon

AdiPolak · x · 2026-11-18

At QCon, OpenAI's Mouli shared how to make Spark upgrades boring using a sheet plus a swarm of agents: claim a job, run a shadow with pinned inputs and private outputs, read the report, then roll out gradually—pausing or debugging when live checks or shadow runs fail, and writing statuses back to the sheet.

The shadowing system runs production and candidate changes (image, code, or config) side by side with isolated shadow outputs, generating comparison reports for validation. A practical blueprint for agent-driven large-scale upgrades.

Original post →

More from coding & agent

coding & agent channel →