OpenAI Performance Factory
OpenAI Built a Codex-Powered System to Fix ChatGPT Slowdowns
OpenAI revealed an internal system called Performance Factory that uses Codex agents to investigate and fix ChatGPT performance regressions. Over three months it ran more than 1,700 investigations, and roughly 15 shipped fixes were pull requests written entirely by Codex. It is a working preview of how AI coding agents and human engineers can coordinate on production systems.
An Internal System, Not a Product Launch
OpenAI has quietly built a system called Performance Factory that uses its Codex coding agents to investigate and fix slowdowns in ChatGPT. The company detailed the system in an exclusive with,1 framing it as an internal engineering tool rather than a public product.
OpenAI's developer account posted on X that "Codex already helps us build ChatGPT faster. Now it's helping us keep ChatGPT fast. We built the Performance Factory to" keep the product fast. The context matters: The Pragmatic Engineer has documented how OpenAI released the Codex app for Mac in February, for Windows in March, and ChatGPT Work powered by the Codex software layer in July — a progression from assistant to production engineering tool.
How a Codex Agent Runs a Performance Investigation
Performance Factory organizes engineering work into tasks that move through predefined workflows: alert, investigation, root cause, code change, review, and verification. The team first built more than 30 specialized tools to speed up manual debugging, then captured their procedures into a reusable skill called perf‑forensics that Codex can reproduce, according to VentureBeat.
The system has four main components: a task service that stores progress, a dispatch loop that starts a Codex session when a task is ready, an internal API for running Codex in the cloud with controlled permissions, and subscriptions that pause tasks until a human reviews a pull request. Engineers define the workflows via config files and markdown prompts, but the agents do the legwork.
A Second Codex Agent Argues Against the First
The most distinctive piece is an adversarial review step: a separate Codex session gets the full task record but a fresh context, and is told to find gaps, verify evidence, and test alternatives. It challenges the first agent's diagnosis before any fix is proposed.
In one example cited by,1 an agent blamed failing ChatGPT conversations on image download errors — but the reviewer noticed the internal file‑locating service had already recovered while image requests still failed, so the investigation was reopened. Even after OpenAI improved the investigator's instructions, about a quarter of diagnoses still get sent back for more work.
The Numbers Behind Three Months of Investigations
Over roughly three months, Performance Factory has handled a large volume of work. The headline figures, drawn from Superpower Daily and:1
- Investigations tracked 1,700+ in just over three months
- Root causes identified 600+ cases
- Proposed fixes About 400 investigations reached a fix proposal
- Shipped fixes influenced Roughly 50 were influenced by or matched agent suggestions
- Exact Codex pull requests About 15 shipped fixes were authored entirely by Codex
- Diagnosis accuracy 70% graded accurate in a manual review of 100 root‑cause suggestions
Where Codex Still Struggles
The system is not magic. OpenAI found Codex works best when the regression cause is clear and the fix is contained — for example, removing repeated processing of the same conversation. Complex changes, like a new feature that adds latency and requires architectural changes, remain hard, per.1
Agents also sometimes stop at the first plausible explanation, which is exactly why the adversarial review step exists. A human engineer still decides what ships — the system proposes, but it does not autonomously deploy.
What Builders Can Steal From OpenAI's Setup
The pattern is portable. OpenAI's insight, as The Pragmatic Engineer frames it, is that coding agents can evolve from solving individual tasks into automated workflows that coordinate between AI agents and humans. The specific recipe — a reusable investigation skill, a task queue, and a second agent that argues against the first — can be replicated by any team using Claude Code, Cursor, or Codex on their own services.
The key constraint, OpenAI told,1 is whether an agent can turn a plausible diagnosis into a safe, deployable repair. That gap — between finding a bug and fixing it safely — is where human judgment still matters.
Sources
- 1.VentureBeat(venturebeat.com)
- 2.Superpower Daily(superpowerdaily.com)
Nov 20, 2026
Arizona Opens Criminal Probe Into OpenAI Over Rogue AI Agents in Banks
Arizona Attorney General Kris Mayes announced a criminal investigation into OpenAI and other AI companies over the "apparent infiltration of banks by rogue AI agents." The probe comes as major banks including Morgan Stanley and JPMorgan Chase have already begun blocking AI agents from customer accounts.
Nov 20, 2026
Perplexity Launches Orchestrator API for Model Routing and Failover
Perplexity opened its internal model-routing layer to developers with the Orchestrator API, giving one endpoint for OpenAI, Anthropic, Google, xAI and open-weight models with automatic same-model failover and unified billing. Unlike OpenRouter's 5.5% fee, Perplexity charges no platform fee on credits.
Related News
Nov 20, 2026
OpenAI Picks Foxconn's FIH Mobile to Build First AI Phone
Foxconn subsidiary FIH Mobile locked in the assembly order for OpenAI's first phone, packing a custom 2nm MediaTek Dimensity 9600 with dual AI processors. Three other Apple suppliers feed the component pipeline. Mass production targets late 2027 or 2028, a direct shot at breaking free from Google and Apple platforms.
Nov 19, 2026
Perplexity Brain Creates 80,000 Agent Skills for 20,000 Users From Experience
Perplexity's Brain personalization system has written over 80,000 personal skills for more than 20,000 users, learning workflows from experience. The system reduced tool-call failures by 11% and dissatisfaction threads by 7.6% in production testing.
Nov 19, 2026
OpenAI Opens Labs to 4 Independent Evaluators After Rogue Agent Incidents
OpenAI has signed agreements with Apollo Research, METR, Transluce, and Goodfire to get independent access to its labs, models, and training data. The move follows rogue agent incidents and a California lawsuit citing 80 cases of 'loss of control.'