OpenAI Introduces Performance Factory, a Codex-based Software Optimization System That Supports High-Speed ChatGPT
OpenAI has developed an internal Codex-based AI system called Performance Factory that automatically detects and suggests fixes for ChatGPT slowdowns after code updates. According to an engineering report, in just three months, the system conducted over 1,700 investigations (an average of 17 per day) and identified the root cause in over 600 cases.
OpenAI has developed an internal Codex-based AI system called Performance Factory that automatically detects and suggests fixes for ChatGPT slowdowns after code updates. According to an engineering report, in just three months, the system conducted over 1,700 investigations (an average of 17 per day) and identified the root cause in over 600 cases.
Why do we need a system?
As AI agents increasingly write and modify code, manual performance testing becomes a development bottleneck, VentureBeat writes . Previously, OpenAI engineers had to manually analyze logs, traces, and code changes using more than 30 specialized tools. To automate this process, the team first created a perf-forensics skill for Codex, and then built a full-fledged task system that independently responds to monitoring alerts.
How the Performance Factory works
The system architecture consists of four key elements.
Task Service: Stores the progress of each incident, metadata, and summary agent findings. Instead of sending huge amounts of raw data, agents record key metrics and queries so that the next agent can replay the inspection if needed.
Dispatching Cycle: Periodically checks tasks and launches Codex sessions to move to the next stage.
Internal API for Cloud Codex: Provides agents with controlled access to code repositories and watchdog systems.
Subscriptions: link tasks to external events — for example, pausing a process until an engineer checks a pull request on GitHub.
"Competitive test" and real results
During the work, the team found that the agents sometimes stopped at the first plausible explanation for the failure. To avoid mistakes, OpenAI introduced adversarial review: a separate Codex session with a “clean” context is tasked with refuting the previous agent’s diagnosis, finding gaps in the evidence, and testing alternative hypotheses. Currently, such a reviewer returns about 25% of the conclusions for revision. A similar check was added for proposed code fixes.
Statistics show that the system already significantly saves time on diagnostics, although the generation of the finished code still needs further refinement:
of over 600 investigations with an established cause, about 400 have reached the stage of proposing a fix;
approximately 50 final production changes relied on AI suggestions;
About 15 fixes were incorporated into the code directly from pull requests created by Codex.
Codex is best at handling local issues (like code serializing the same conversation multiple times while processing a request), while complex architectural changes are better handled by humans. After an engineer deploys a fix, another agent automatically checks to see if the issue is resolved.
Business plans and conclusions
For teams implementing AI agents, OpenAI engineers have identified several practical rules: group reports for humans (to avoid “AI fatigue” from excess generated text), pass reproducible queries between agents instead of raw logs, and run separate agents to monitor errors in the AI tools themselves.
In the future, OpenAI plans to move Performance Factory from reacting to failures to proactive optimization: agents will compare processor and memory profiles between releases, optimize autoscaling settings, and in the future will be able to independently deploy low-risk fixes without human intervention.
Як нейромережі бачать вільну та незалежну Україну? Тест dev.ua
Нейронні мережі для генерації зображень бачать світ по-своєму, їхню логіку зрозуміти часом зовсім неможливо. Але таки хочеться. На честь Дня Незалежності України редакція dev.ua вирішила провести невеликий експеримент.
Ми задали чотирьом різним нейронним мережам п’ять однакових запитів: «прапор України», «День Незалежності України», «український Крим», «перемога України» та «українці». Отриманими результатами ми ділимося з вами нижче.
У TikTok тепер можна генерувати фон за допомогою нейромережі. Ми протестували її та ділимося результатами
У TikTok з’явилася нова функція «Розумний фон». З її допомогою як фон для тіктоків можна підставляти згенеровані нейромережею зображення. Редакція dev.ua протестувала цю технологію і ділиться своїми враженнями.
“The decision always remains with the person”: Ukrainian developers have created the AI platform Talkruit, which takes over the entire hiring funnel — from screening to checking test scores
Two Ukrainian developers, Marko Holyak and Lubomyr Pytulia, launched the AI startup Talkruit, which combines the entire IT hiring funnel in one window — from synchronizing job board responses to conversations with a voice agent and running test cases in a sandbox.
Co-founder of the startup, Lyubomyr Pytulya, told dev.ua how the pet project grew into a commercial product, how much development costs, how monetization works, and what the competition is like in the HR technology market.
Have important news to share? Message our Telegram bot
Key events and useful links in our Telegram channel