What's this about?
A Latent Space episode where shloked reconstructs how ChatGPT Work — OpenAI's Codex harness extended from coding to cloud and general knowledge work — implements Memory, Proactivity, and more. It launched July 9 and crossed 10M users three weeks later.
What exactly is a "harness" (again)?
The engineering scaffolding around a model that makes it a useful agent — how it manages memory, decides when to act proactively, calls tools, and handles context. shloked's series reverse-engineers how each frontier lab builds theirs; this is his deepest ChatGPT dive.
Why does it matter?
Harness engineering is a recurring theme this week (the source-code reading lists, everyone's multi-model workflows). A careful teardown of the harness behind a product that hit 10M users in three weeks is a rare concrete look at what "good harness design" actually consists of.