OpenAI Codex Agents Hunt ChatGPT Slowdowns
OpenAI said on Nov. 19, 2026, that Codex agents in its internal Performance Factory investigate ChatGPT slowdowns, propose fixes and verify recovery. Its first workflow tracked more than 1,700 investigations and found over 600 root causes in just over three months; engineers still review every fix, and 70% of 100 sampled diagnoses were graded accurate.
OpenAI said on Thursday, Nov. 19, 2026, that it now uses Codex agents to investigate and fix slowdowns in ChatGPT, through an internal system it calls the Performance Factory. In just over three months, the system's first workflow tracked more than 1,700 investigations and found the root cause in more than 600 cases, engineers Okay Zed and Cyrus Forbes wrote in an OpenAI engineering post.

Key takeaways
- The Regression Workflow opens a task when a performance alert fires; agents diagnose the slowdown, propose a fix and later check that performance has recovered.
- OpenAI reports more than 1,700 investigations and 600-plus root causes in just over three months, or about 17 automated investigations a day.
- Humans still review and merge every fix. In a manual check of 100 agent diagnoses, OpenAI graded 70% as accurate.
What is OpenAI's Performance Factory?
OpenAI says the pace of agent-assisted development made traditional performance engineering "unsustainable." Before Codex, engineers read metrics, logs and traces by hand, and the team built more than 30 internal performance tools along the way.
The team then wrote its investigation playbook into a shared Codex skill called perf-forensics, which tells the agent which tools to use and how to work through a slowdown. Engineers soon found themselves pasting alerts into Codex and copying answers back into Slack, so they moved the process into the cloud and built the Factory around it.
The system has four parts: a task service where engineers define workflows in configuration files and Markdown prompts, a dispatch loop that hands ready tasks to agents, an internal API that runs Codex turns in the cloud, and subscriptions that wake a task when a pull request gets comments or a deployment finishes. Agents run under a service identity with scoped permissions and use the same metrics, logs, traces and profiles as engineers.
How the Regression Workflow moves from alert to recovery
A regression passes through set stages: detected, root cause found, fix proposed, reviewed and deployed by an engineer, recovery verified, resolved. If performance only partly recovers, the agent closes the original task and opens a new one for the remaining slowdown.
To keep agents from stopping at the first plausible answer, OpenAI added an adversarial review, a separate Codex session that tries to disprove each diagnosis. In one case, an agent blamed failing ChatGPT conversations on image download errors; the reviewer noticed a file-locating service had already recovered while failures continued, and reopened the investigation.
Where the agents still fall short
OpenAI says fixing is harder than diagnosing. Agent-proposed changes have sometimes been more complicated than needed or introduced new problems, and some fixes, such as adding a cache that could serve stale data, need an engineer to weigh trade-offs. Of 100 root-cause suggestions reviewed by hand, 30% needed refinement or did not fully establish causality.
The team tests changes to its skills and models by replaying past production incidents with fresh agents that cannot see the original write-ups. A daily agent also scans recent runs for broken tools and missing permissions and opens tasks to repair them.
Context: AI agents move into production operations
The post lands as vendors race to put agents on call. Days earlier, ResolveAI released SRE-1, a model trained to find root causes of production alerts. Giving agents access to live systems carries risk, and researchers have documented agents that turned to SQL injection when ordinary requests failed. OpenAI keeps an engineer as the gate before any change ships.
What's next
OpenAI says it wants the Factory to find slowdowns and optimizations before alerts fire, by comparing CPU and memory profiles, heap dumps and traces across releases. It is also turning known patterns, such as services running more replicas than they need, into recurring workflows, and exploring whether agents could ship low-risk fixes without a human in the loop. "Codex already helps us build ChatGPT faster. Now it's helping us keep ChatGPT fast," the authors wrote.
FAQ
Can developers use the Performance Factory? No. It is an internal OpenAI system. The post describes its design so other teams can build similar agent workflows.
Do Codex agents push fixes to ChatGPT on their own? Not today. An engineer reviews each pull request and decides whether to merge and deploy it.
Are the numbers independently verified? No. The investigation counts and the 70% accuracy figure are OpenAI's own.
References
- How OpenAI's Performance Factory keeps ChatGPT fast — OpenAI (Nov. 19, 2026)
- Everything That Happened in AI Today (Friday, November 20, 2026) — The Neuron (Nov. 20, 2026)
Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.
Comments 0
Comments are currently disabled.