OpenAI’s Codex agents tracked more than 1,700 investigations into ChatGPT slowdowns in just over three months and root caused over 600 cases, OpenAI said on 19 November.
That is 17 automated investigations every day, the company says, run by a system it calls the Performance Factory.
Business Pill · Lead Time Explained
An evergreen Business Engineer explainer on lead time. It teaches the general idea only and does not discuss OpenAI or Codex.
The key insight: OpenAI says its Codex agents ran more than 1,700 ChatGPT slowdown investigations in just over three months, about 17 a day, with engineers approving each fix.
What OpenAI Built
OpenAI says “the staggering pace of agentic development has made traditional approaches to performance engineering unsustainable”. Its answer is a cloud system where Codex agents find and fix regressions.
Engineers write workflows and Markdown instructions; agents do the steps. The first one, the Regression Workflow, opens a task when a performance alert fires.
An agent investigates, records its findings and proposes a fix. An engineer reviews and deploys it, and another agent then checks whether performance returned to baseline.

The Four Parts
The system has a task tracker with configurable workflows, a dispatch loop that assigns ready work to agents, an internal API that runs Codex in the cloud, and subscriptions that wake agents on events such as a GitHub review comment or a finished deployment.
The agents run under a service identity, so OpenAI can scope their permissions and see which actions came from its performance agents, the post says.

How Accurate the Agents Are
In a manual review of 100 root-cause suggestions, OpenAI graded 70% as accurate. The other 30% needed further refinement or did not fully establish causality.
Early agents sometimes stopped at the first plausible explanation. OpenAI added an adversarial review: a separate Codex session that tries to disprove each diagnosis, and later each proposed fix.
OpenAI says “getting fixes right has been harder than finding the cause”. Some agent fixes were more complicated than necessary or introduced new issues.
One Case From the Factory
OpenAI’s sample report, REG-482, covers 27 to 28 October. A change to worker registration raised median worker startup time from 20.6 seconds to 45-50 seconds, and a rollback restored the baseline on 28 October.
In matched traces, the same startup phase took 1.18 seconds on the old build and 46.34 seconds on the new one, per the report. Its proposed fix was still a draft.

What Comes Next, per OpenAI
OpenAI says it is exploring whether agents can identify low-risk changes and “eventually ship those fixes without a human in the loop”. Today, it says, the bottleneck may be waiting for an engineer to review and merge the fix.
It is also working on workflows that find slowdowns even when no alert fires, and wants to run known patterns across all its services, such as its existing service-rightsizing skill, which spots autoscaling settings that keep more replicas running than a service needs.
The Structural Read
As we read it, OpenAI is using its own operations as a showcase for agent workflows: alerts in, diagnoses and pull requests out, humans at the merge step.
As we read the 70% figure, OpenAI is publishing an error rate alongside the volume, which is why it built a second agent to argue with the first.
OpenAI, 19 Nov 2026
“the staggering pace of agentic development has made traditional approaches to performance engineering unsustainable”
Three Implications
FOR ENGINEERING TEAMS OpenAI describes a pattern others can copy: tasks with explicit states, agents per step, an adversarial reviewer and an engineer at the merge.
FOR CODEX The post shows Codex running unattended in the cloud on production telemetry, not only as a coding assistant.
FOR RELIABILITY OpenAI’s own review graded 70% of 100 root-cause suggestions accurate, so human review stays in the loop for now.
What Is Not Established
OpenAI gave no latency gains for ChatGPT users, no cost figures and no count of fixes merged or shipped. The 70% accuracy grade is OpenAI’s own review of 100 suggestions, not an outside test.
The Bottom Line
OpenAI now runs ChatGPT performance work through Codex agents: more than 1,700 investigations in just over three months, over 600 root caused, with engineers still approving every fix, the post says.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Every fact in this piece is from OpenAI’s 19 November engineering post, including its example task report. The share of investigations root caused (about 35%) is our arithmetic on OpenAI’s figures. We have not seen outside measurements of the system.
Sources: OpenAI, ‘How OpenAI’s Performance Factory keeps ChatGPT fast’, 19 Nov 2026









