OpenAI engineer Rodrigo Roim says Codex now finds dead code through Slack and production data, and his team merged over 1,400 cleanup PRs
Software maintenance is changing. A few months ago, I had to tell Codex exactly where to look for dead code. Now it can find candidates in Slack, check them against production telemetry, and draft cleanup PRs. Here’s how we do this at OpenAI.
Context
Maintenance is a large hidden cost in software, and OpenAI's numbers suggest coding agents can take on cleanup work that teams rarely find time for.
This is a first hand post from Rodrigo Roim, an applied AI engineer at OpenAI, linking a blog post on OpenAI's developer site. The results describe OpenAI's own codebase and are not independently verified.
Deep dive
On October 16, Rodrigo Roim posted that software maintenance is changing. He wrote that a few months ago he had to tell Codex exactly where to look for dead code, but now it can find candidates in Slack, check them against production telemetry and draft cleanup PRs. OpenAI Developers quoted the post later that day.
In the linked blog post, Roim says he works on OpenAI's Core Experimentation team. He writes that after GPT 6 Astra handled a short, vague prompt well, five engineers ran variations of it across repositories and merged more than 1,400 pull requests, deleting over 800,000 lines of code, mostly part time over two weeks. One service's container image shrank 48%, from 3.31GB to 1.72GB.
He also describes using Astra to find stale experiments and feature flags, including one flag whose removal saved an estimated 0.42% of CPU in a service. In a comparison on the same task, GPT 6.1 Sol Ultra produced 92% as many findings as GPT 6 Astra Ultra at 14% of the API cost.
Roim notes that some proposals added more complexity than they were worth, so measurement and review still mattered.
What to expect next
Watch whether other engineering teams report similar cleanup results with Codex or rival agents, and whether OpenAI turns these prompts into built in Codex features.
Also watch for mistakes from automated code removal, which Roim says caused incidents in the past even with older tools. Cost comparisons between models on real maintenance work, like the one in the post, will help teams decide which model to use.