OpenAI’s Performance Factory: Codex agents now investigate ChatGPT’s own performance regressions — constrained agents doing production engineering
What this story is about
SXF Brief’s engineering file: OpenAI explains how Codex-backed agent workflows help investigate ChatGPT performance regressions and automate repeatable production engineering checks — agents turned inward, on the operator’s own systems, with bounded scope. It is the constructive counterpart to the week’s agent-escape litigation: the same class of tooling, run inside a constrained harness against a defined target, producing checkable work. The desk’s interest is the pattern, which generalizes — an agent with a narrow mandate, a real system, and a human reviewing the diff is the deployment shape that keeps working, in a week when the unconstrained version is being sued in two states.
Why it matters
Agents investigating their own platform's regressions, inside a harness, is the deployment shape that keeps surviving contact with reality -- the constructive half of a week otherwise spent on agent litigation.
About the source
SXF Brief — AI-news briefing
Independent curated AI brief (SXF) with an explicit verified-source-facts format -- facts are checked against named primary sources and the brief says what it could not verify. Strong on launch-day AI infrastructure and agent tooling.