OpenAI on Nov. 19 described an internal system that uses Codex agents to investigate ChatGPT slowdowns and propose fixes, according to a company blog post1.
The announcement offers a concrete example of how OpenAI says it is moving coding agents into persistent operational workflows, while also showing that automated diagnosis is advancing faster than fully automated repair, by the company's own figures.
OpenAI said the system, called Performance Factory, runs in the cloud and automates recurring performance-engineering work by turning alerts into tracked tasks that can move from investigation to root-cause analysis, code changes and post-deployment verification, according to the company blog post1. The announcement includes no independent test results.
The company said its first workflow, called the Regression Workflow, tracked more than 1,700 investigations and root-caused more than 600 cases in just over three months, or 17 automated investigations a day, according to the company blog post1. Those figures are OpenAI's own account of internal use, not a comparative benchmark.
System components and evolution
According to the company blog post1, Performance Factory has four components: a task tracker with configurable workflows, a dispatch loop that assigns work to agents, an internal API for running Codex in the cloud, and subscriptions that wake agents when an external event such as a code-review comment or deployment completes.
OpenAI said engineers define workflow states and prompts in configuration files and Markdown, while agents run inside OpenAI's network with access to code repositories and observability tools including metrics, logs, traces, CPU profiles and memory snapshots, according to the company blog post1. The company said agents operate under service identities so permissions can be scoped and actions traced.
The system grew out of earlier stages in which engineers first investigated regressions manually, then used Codex as an assisted debugging tool, then encoded recurring procedures in a shared skill called perf-forensics, according to the company blog post1. OpenAI said its performance team had accumulated more than 30 performance tools before building the broader workflow system.
Accuracy and repair limits
OpenAI said agents initially tended to stop at the first plausible explanation for a slowdown, so it added an "adversarial review" in which a separate Codex session tries to disprove the proposed diagnosis, according to the company blog post1. In one example, OpenAI said a reviewing agent reopened an investigation after finding that an earlier diagnosis about image-download failures did not explain why requests were still failing.
In a manual review of 100 root-cause suggestions, OpenAI said it graded 70% as accurate, with 30% needing further refinement or not fully establishing causality, according to the company blog post1. The post does not provide the sampling method, model configuration or an external baseline for that review.
Secondary coverage by VentureBeat2 reported additional OpenAI-supplied figures not included in the company post: around 400 of the 600-plus root-caused investigations progressed to a proposed fix, about 50 shipped fixes were influenced by or matched agent suggestions, and around 15 shipped fixes were exact pull requests authored by Codex. Those figures also come from OpenAI's performance team and were not independently verified.
That gap between investigation volume and shipped fixes is central to what the numbers show. OpenAI's own description indicates the system is further along at triage, diagnosis and workflow management than at autonomously generating production-ready repairs, especially for broader architectural changes, according to the company blog post1 and OpenAI comments reported by VentureBeat2.
Human review remains
OpenAI said engineers still review proposed pull requests and decide whether to merge them, according to the company blog post1. The company said tasks can pause for review comments or deployments and resume when matching events arrive, rather than keeping an agent active while waiting.
To make that handoff usable, OpenAI said each task includes a generated summary of the regression, its impact, the suspected root cause and the proposed fix, along with artifacts such as metrics series, flame graphs and traces, according to the company blog post1.
OpenAI also said it has completed a workflow evaluation suite that it plans to use to improve agent-generated fixes, according to figures attributed to the performance team in VentureBeat2. The company post says it is also exploring additional workflows, including spotting optimizations that could trigger new tasks in other workflows, according to the company blog post1.