2026-11-19Gunner Technology

What OpenAI's Performance Factory Actually Automates

OpenAI's Performance Factory handles slowdowns by turning alerts into tracked tasks that agents investigate, challenge, propose fixes for and check after deployment. Engineers still decide whether to merge the proposed code. The useful part is how the work keeps moving between those steps without somebody carrying every finding from one window to another.

What starts an investigation?

A configured performance alert creates a task for agents to investigate. A performance regression means an application has become slower or more resource hungry after a change. According to VentureBeat's report on Performance Factory, OpenAI uses this internal system to investigate regressions affecting ChatGPT.
The team got there in stages. Engineers first used specialized investigation tools themselves, then had Codex, OpenAI's coding agent, operate those tools. They captured their procedure in a reusable skill called perf-forensics. Eventually, alerts could start cloud sessions directly, removing the job of copying a warning into the agent and carrying its answer back to chat.
That still left a pile of diagnoses for people to organize. The next step was to make each regression a durable task with a route through investigation, review, repair and verification. The report describes an internal operating system for this work, rather than a deployment recipe a buyer can simply install.

How does the work survive between agents?

The task record carries the findings and progress between sessions. It holds written notes alongside structured facts, including when the slowdown began and when its cause was identified. Agents leave relevant measurements and the queries used to obtain them, so the next worker can inspect the evidence without receiving a mountain of raw profiling data.
A task service stores the record. A dispatch loop checks for work ready to advance and starts an appropriate Codex session. An internal interface provides cloud execution with controlled access to repositories and monitoring tools. Event subscriptions let the workflow wait for something outside the session, such as an engineer's feedback on a proposed code change.
This is a concrete example of why the workflow has to survive the agent. The question is no longer whether a particular conversation remembers the investigation. The task tells the next session what happened, where the evidence lives and what step comes next.

Who checks whether the diagnosis holds up?

A separate Codex session challenges the investigator's diagnosis before the workflow accepts it. The article calls this adversarial review: a reviewer deliberately looks for unsupported claims and competing explanations. It starts with fresh context, receives the task record and tests whether the proposed cause accounts for the observed failure.
The report describes an investigation where a reviewer noticed that a service had recovered while the failures being explained continued. That mismatch sent the task back. It is a useful example because the reviewer challenged the timeline, rather than merely deciding that the original explanation sounded plausible.
The article says OpenAI also added this review to proposed fixes. Keep the limit clear: a second agent does not automatically provide independent proof. Gunner Technology recommends preserving the underlying evidence and preventing the builder from changing the acceptance conditions. A fresh session helps challenge a story; enforceable boundaries decide what it may approve.

How much repair is actually automated?

Investigation is further along than automatic repair in the system the article describes. VentureBeat reports that roughly 50 shipped fixes were influenced by or matched agent suggestions, while about 15 were the exact agent-authored pull requests. A pull request is a proposed code change submitted for review. Those are different outcomes and should stay separate when assessing the results.
An engineer reviews the proposed change and decides whether to merge it. The report says contained fixes with clear causes work better than broader architectural changes. It also describes the danger of an apparently useful optimization: adding a cache can reduce waiting while leaving users with outdated data.
These are results reported by OpenAI through the article, not an independent comparison or a forecast for your repository. Proactive optimization and deploying some low-risk fixes without human approval appear as future directions. They should not be presented as capabilities already operating in the reported workflow.

When is the slowdown actually resolved?

The workflow checks production after deployment and returns an unresolved problem to investigation. That last step matters because accepting code does not establish that the affected requests recovered. The task remains the thread connecting the original alert, the proposed explanation, the deployed change and what happened afterward.
The system also groups related incidents and presents concise reports with measurements and graphs. That gives engineers a way to judge impact without rebuilding each investigation from agent output. Automating the investigation would lose much of its value if understanding its results became another full-time sorting job.
For a buyer, Gunner Technology recommends starting with one service and one recognizable class of slowdown. Define the affected requests, the evidence required for recovery and who approves changes. Keep the work and access in your own environment. As with any change that must survive production, carry the task past the merge. Let agents do the tracing and routing; keep people responsible for intended behavior and release decisions.

Quick answers

Do we need to automate fixes to benefit from this approach?

No. Automating alert intake, evidence gathering and task tracking can remove manual work while engineers continue to approve every code change.

What if an investigation query stops working?

Record that the evidence is unavailable and keep the diagnosis unverified. Restoring access or preserving the relevant data is part of completing the investigation.

Should agents receive unrestricted production access?

No. Give each stage only the access its job requires, and keep deployment permission separate from permission to inspect monitoring data.