ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
TopicDisconfirmAI EditorialUpdated 2026-11-22 15:57

OpenAI Performance Factory

3 steps1 pieces3 Sources

OpenAI has deployed the "Performance Factory," an internal autonomous system powered by Codex that automates the investigation and mitigation of performance regressions in ChatGPT. The system addresses a critical bottleneck: as AI agents accelerate code changes, manual performance review becomes unsustainable. In just over three months, its first workflow, the Regression Workflow, tracked more than 1,700 investigations and identified root causes for over 600 cases, averaging 17 automated investigations per day.

The factory operates through four components: a task service with configurable workflows, a dispatch loop assigning work to agents, cloud-based Codex execution, and event subscriptions that wake agents when conditions change (e.g., GitHub reviews or deployments). A key innovation is the use of independent adversarial agents to rebut initial diagnoses, improving accuracy. However, while diagnosis is automated, fixes remain harder to implement; only about 15 shipped fixes were exact pull requests authored by Codex, with engineers still reviewing and merging most changes. OpenAI reports a 70% root-cause accuracy rate in manual assessments, but emphasizes that architectural tradeoffs and final deployment decisions require human judgment.

How it got here

  1. 2026-11-19

    November 19, 2026: OpenAI engineers Okay Zed and Cyrus Forbes publish an engineering post detailing the Performance Factory, describing its evolution from manual tools to automated Codex workflows.

    openai.com ↗
  2. 2026-11-19

    November 19, 2026: VentureBeat releases an exclusive report on the system, highlighting metrics such as 1,700+ investigations and 600+ root causes identified in three months.

    venturebeat.com ↗
Sign in / Sign up
ReadingResearchRadarInvestment framework
  • 2026-11-20

    November 20, 2026: Additional analyses emerge, noting that while 400 cases progressed to proposed fixes, only ~50 shipped fixes matched agent suggestions, and ~15 were direct Codex PRs.

    ai-primer.com ↗
  • What is still unsettled

    • Can the system eventually deploy low-risk fixes without human approval?
    • How will the 30% of inaccurate or incomplete diagnoses be further reduced?
    • What are the long-term impacts on engineering team structure as automation expands?

    Sources

    • openai.comOpenAI Engineering Blog↗
    • venturebeat.comVentureBeat Exclusive Report↗
    • ai-primer.comAI Primer Analysis↗

    Coverage in this topic

    • OpenAI Performance Factory: Independent Agent Rebuttals Improve Root-Cause Accuracy2026-11-20News

      OpenAI engineers Okay Zed and Cyrus Forbes detail their Performance Factory system, explaining how early agents sometimes stopped at the first plausible explanation to provide incorrect diagnoses, and how introducing an independent agent specifically tasked with rebutting those diagnoses intercepts these errors—achieving a current root-cause accuracy of 70 out of 100 cases, though fixes remain harder to implement than diagnoses.