Field noteNov 20, 20265 min read1 source

Codex already helps us build ChatGPT faster.

In one minute

OpenAI says its Performance Factory uses Codex agents to investigate ChatGPT performance regressions, propose fixes for engineer review, and verify recovery after deployment. The workflow has tracked more than 1,700 investigations in just over three months. Its design is a concrete example of turning recurring operational work into stateful, event-driven agent workflows with explicit human checkpoints.

What happened

OpenAI’s November 19, 2026 engineering announcement describes the Performance Factory as a cloud-based system for investigating and resolving performance problems. Its first workflow, the Regression Workflow, begins when a configured performance alert fires. The system opens a task, and agents move it through stages: needs investigation, root caused, fix proposed, and resolved. In just over three months, the workflow tracked more than 1,700 investigations and root-caused more than 600 cases—about 17 automated investigations a day.

The system has four parts: a task service with configurable workflows, a dispatch loop, an internal API that runs Codex in the cloud, and subscriptions that wake waiting tasks when events occur. Engineers define task states, allowed transitions, and steps requiring human involvement in configuration files, with Markdown prompts for each agent step. The dispatch loop periodically finds ready tasks and starts a Codex turn for the current step. Tasks retain their status and findings outside any single agent’s context, so work can resume after review or deployment.

OpenAI says Codex runs in an environment inside its network, with access to repositories and observability tools including metrics, logs, traces, CPU profiles, and memory snapshots. Agents operate under a service identity, allowing permissions to be scoped and actions attributed to performance agents. A task can wait for events such as a GitHub pull-request review comment or failed CI check; a matching event makes it ready for dispatch again.

For diagnosis, the perf-forensics skill guides agents to identify when a slowdown began, which requests were affected, and whether deployments or configuration changes explain it. Agents compare request groups, inspect deployments and feature-flag changes, and use logs or traces to test hypotheses. OpenAI added a separate adversarial Codex session to challenge proposed causes and check counterfactuals. In a manual review of 100 root-cause suggestions, the company graded 70% accurate; 30% needed refinement or did not fully establish causality.

After diagnosis, Codex can propose a fix and open a pull request, but an engineer reviews it and decides whether to merge. Once a fix is deployed, an agent checks whether performance returned to baseline. Partial recovery leads to a new task for the remaining slowdown; no improvement sends the issue back for investigation. OpenAI also describes backtesting against past incidents, separate evaluation of engineer-facing summaries, and a daily agent check for tool failures and access errors.

What the archive adds

New Runtime's earlier coverage includes an OpenAI retirement affecting ChatGPT and Codex, while leaving API access unaffected (November 1, 2026), and a report of an October degradation affecting Codex, ChatGPT, and APIs (October 28, 2026). Those notes concern product access and reliability incidents; this is a new event, focused on OpenAI’s internal operational system for finding and addressing performance regressions.

Trend signal

This case supports AI-native organizations move toward review and orchestration: agents handle investigation and proposed changes, while engineers review evidence, weigh tradeoffs, and decide what ships. It also illustrates how work can be routed through explicit states rather than handed to a single agent with an open-ended mandate.

Why now

OpenAI says the pace of agentic development has made traditional performance engineering unsustainable. Its factory is intended to help service teams keep up by automating recurring investigation and follow-through, and the company is exploring proactive detection beyond alerts.

Why it matters

The useful unit here is not a clever debugging prompt but an operational loop with durable task state, scoped access, event-driven resumption, review, and verification. That structure addresses several failure modes that appear when agents are asked to handle production work: losing context between steps, stopping at a plausible but wrong diagnosis, proposing a fix that creates new problems, and leaving teams to reconstruct a wall of agent output.

Teams can borrow the workflow pattern without copying OpenAI’s internal system: turn a recurring alert into a task with named states; preserve evidence and findings across agent sessions; add a distinct challenge step for causal claims; require a human to approve risky changes; then verify the result against a baseline. OpenAI’s own 70% accuracy figure is also a reminder that agent-generated root causes need evaluation and review, not blind trust.

Limits and uncertainty

The announcement describes OpenAI’s internal implementation and its own reported activity; the figures are not an independent evaluation. The 70% accuracy result comes from a manual review of 100 root-cause suggestions, and OpenAI says the remaining 30% needed refinement or did not fully establish causality. It does not provide a comparison with the prior process’s time or cost, or a measured rate for fixes accepted, merged, or causing regressions. The company says it is exploring more proactive optimization and greater autonomy, but those are future work, not capabilities established by this account.

Skill to build now

For one recurring production alert, define a small workflow this week: record the signal, ask an agent to identify when and where it changed, require a separate review to test the leading cause against a counterfactual, and keep deployment approval with an engineer. Save findings in a task record and make post-deployment verification a required step. OpenAI’s process suggests evaluating the investigation and the handoff summary separately, rather than judging only whether the agent produced a plausible answer.

Reader takeaway

OpenAI’s account shows a practical way to make agents part of production operations: explicit task states, persistent findings, event-triggered continuation, adversarial review, human approval, and verification after deployment. The company’s reported volume is notable, but the more reusable lesson is that reliability work needs a workflow around the agent—and a way to test whether its conclusions hold up.

Evidence / sources

  1. 1openai.com/index/how-openai-performance-factory-keeps-chatgpt-fast
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicCodexExplore the codex topic hub.
  2. 02archiveField NotesOpen the latest editorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract