OpenAI engineering manager Thibault Sottiaux wrote that Codex now pretty much builds itself, with humans as the limiting factor in the loop (X), as OpenAI launched a Codex app for macOS built as a command center for running many agents at once. It escalates the December report that Codex was babysitting its own training runs from an operational detail to a description of the whole product, and mirrors the pattern at Anthropic where Claude Code’s creator stopped opening an IDE. Three days later OpenAI made the claim official, shipping GPT-5.3-Codex as a model instrumental in creating itself; in the same issue, Altman said asking Codex for ideas made him feel useless.
Development
A Codex manager says the product now builds itself
The recursion stated from inside the team that runs it: the engineers building the flagship coding agent name themselves, not the model, as the bottleneck.
Claim
A Codex engineering manager said Codex now pretty much builds itself and identified humans as the limiting factor in the loop, as OpenAI launched a macOS Codex app to serve as a command center for agents.
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to February 3, 2026 (theinnermostloop.substack.com)
- Thibault Sottiaux on X: Codex now pretty much builds itself (x.com)
- Introducing the Codex app (openai.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Knowledge graphopen in browser →
About
Related to
Reported in
Supersedes
Referenced by
- A model finds 500 zero-days, some decades old
- OpenAI ships a model it says was instrumental in creating itself
- Effectively all of Anthropic's product code is written by Claude
- OpenAI begins monitoring its own coding agents for misalignment
- A model tops the leaderboard for writing the GPU kernels it runs on
- Users metaprompt an agent to write its own objective file
- A coding agent generates 99.8% of its own weekly output tokens
- A lab touts an 80% price cut, a billion users, and agents writing 99.8% of tokens
- Welcome to February 3, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.