Wednesday, 18 November 2026

Watch Codex turn Grafana telemetry into quick Kubernetes fixes

OpenAI's demo walks through three on-call emergencies, from failing checkouts to crashing containers, with Codex reading the evidence and engineers signing off on every fix.

Screen showing a Kubernetes rollout status

The short version

  • OpenAI posted a video showing Codex helping engineers investigate live service problems using Grafana dashboards and Kubernetes.
  • The three demos cover a checkout failure, a rollout that triggers out-of-memory restarts and a Codex Security finding about a missing resource limit.
  • In each case Codex gathers telemetry, release history and code, while an engineer reviews the proposed change and checks the result.
  • Since the video, OpenAI has added scheduled security scans to Codex and launched Daybreak, a set of agents for finding and patching flaws.
Quick read · 1 min

OpenAI has shown how its Codex coding agent can help engineers fix live services, not just write code. A short video walks through three emergencies: a checkout failure spotted on a Grafana dashboard, a Kubernetes update that keeps crashing from running out of memory, and a Codex Security finding that a missing resource limit could take a service down.

In each case Codex collects the clues from monitoring data, recent releases and the code, then proposes a fix that an engineer reviews and checks.

For most people, the payoff is shorter outages at the apps and shops they use. OpenAI has kept building in this direction, adding scheduled security scans to Codex in September and launching Daybreak, a set of agents that find and patch security flaws, in November.

It is 3 a.m., checkout errors are climbing and you are the engineer on call. That is the scene OpenAI opens with in a short YouTube video showing how its Codex coding agent can help debug a live service, not just write new code.

Most people will never touch Codex. But every time an online store fails at checkout or an app goes down for an hour, someone is digging through dashboards and logs to find out why. If AI tools can shorten that hunt, outages get shorter for everyone, and the way companies keep websites and apps running starts to change.

01

What Codex production debugging looks like in the video

The video, hosted by Tony Loehr and Anke Hao, runs through three production-monitoring demos. Production means the live system real customers use, as opposed to a test copy. In each one, Codex pulls together three kinds of evidence: telemetry (the stream of measurements a running service sends out), release history (what changed and when) and the code itself.

The point OpenAI stresses is that the agent does not act alone. In every demo, an engineer reviews the change Codex proposes and then checks whether it actually fixed the problem. Codex speeds up the detective work, and a person still decides what ships.

Code review UI highlighting a security finding
02

Demo one: a checkout failure in Grafana

The first demo starts with a failing checkout, the most expensive kind of bug for any online shop. The problem shows up in Grafana, a popular tool that turns a service’s metrics and logs into charts and alerts on a dashboard.

Rather than an engineer clicking through panel after panel, Codex follows the trail: what the dashboard shows, which recent release might be involved and where in the code the failure starts. For a tired person at 3 a.m., having those threads collected in one place is the main appeal.

That matters because checkout bugs cost money every minute they last. Shoppers who hit an error usually do not wait around; they close the tab or buy somewhere else. Cutting the time between the first alert and the first real clue is where an assistant like this could earn its keep.

Play video: Production Monitoring with Codex: Grafana, Kubernetes, & Security
03

Demo two: a Kubernetes rollout that runs out of memory

The second demo looks at a Kubernetes rollout. Kubernetes is widely used software that runs apps inside containers (small, self-contained packages) spread across many servers, and a rollout is the process of replacing the old version of an app with a new one.

In this case the new version keeps hitting out-of-memory restarts, meaning containers use more memory than they are allowed, get killed and start again. Codex investigates which change caused the extra memory use, so the team can decide whether to roll back or adjust the setup.

Computer chip
04

Demo three: when a security finding becomes an outage risk

The third demo connects two jobs that are often handled by different teams. A finding from Codex Security, OpenAI’s tool for spotting weaknesses in code, points to a missing resource limit, a cap on how much memory or processing power a service may use.

That sounds like a security footnote, but without a limit, one misbehaving service can starve its neighbors and take the whole system down. The demo shows how the same finding matters for availability, which is the plain question of whether the service stays up.

05

What this means for the apps you use

For ordinary users, the effect is indirect but real. Faster diagnosis can mean shorter outages at the shops, banks and streaming services you rely on. The human review step in every demo also matters: an AI suggesting a fix is very different from an AI pushing changes into a live system unchecked.

There are reasons for caution. AI coding tools are writing more and more software, and experts warn that more code can mean more security holes, with too few people to check it all. Tools like this one only help if engineers keep reading what the agent proposes rather than approving it on autopilot.

06

How Codex has grown since the video

The demo is part of a bigger push. Codex passed five million weekly users in June, and OpenAI has been adding operations and security features around it. Codex Security began inside the company under the name Aardvark and launched as a research preview in March. OpenAI’s own count credits the tool with helping to repair more than 3,000 critical flaws by April, a figure no outside group has checked.

At its DevDay event in late September, OpenAI added Codex Security Cloud, which can scan whole GitHub repositories on demand, on a schedule or whenever new code arrives, then prepare fixes even while a developer’s laptop is closed. It also gave Codex reusable cloud workspaces and an automatic code review option.

In November, OpenAI went further with Daybreak, which pairs Red Team agents that probe web apps for weaknesses with Defense Factory agents that trace problems through the code and draft patches. Engineers still review and test those patches before they are merged.

07

Quick answers

Can Codex fix production problems on its own?

Not in OpenAI’s demos. Codex gathers evidence and proposes a change, and an engineer reviews it and checks the result before anything goes live.

What tools does Codex work with in the video?

Grafana for dashboards and alerts, Kubernetes for running and updating apps, and Codex Security for spotting risky gaps such as a missing resource limit.

Who is this for?

On-call engineers and operations teams who keep websites and apps running, especially those already using Grafana and Kubernetes.

You're reading the quick version.