According to Ars Technica, OpenAI’s product lead for Codex, Alexander Embiricos, revealed that the “vast majority” of the AI coding agent is now built by Codex itself, creating a recursive self-improvement loop. The command-line version of Codex, launched alongside GPT-5 in August 2025, caused external developer usage to jump 20 times, and a specialized GPT-5 Codex model followed in September. Internally, OpenAI engineers used the tool to build the Sora Android app from scratch with just four engineers in a record 18 days, shipping it to the app store in 28 days total. Employees now treat Codex as a “teammate,” assigning it tasks through project management tools like Linear and communicating with it in Slack, where it can create pull requests from feedback. The system even monitors its own training runs and processes user feedback to decide what to build next.
The recursive loop
Here’s the thing that’s both fascinating and a little mind-bending. We’re not just talking about an AI that writes code for other projects. We’re talking about an AI that is actively writing the code for the system that *makes* the AI. Embiricos described Codex writing the research harness for its own training runs and monitoring those runs. That’s a level of recursion that feels like something out of a sci-fi novel, but it has a very real historical precedent in computing. Think about it: engineers used the first chips to build computers that could run design software, which then let them design chips too complex for humans. Codex building Codex is the software version of that same bootstrapping cycle. It’s a tool building a better tool, ad infinitum.
AI as teammate, not just tool
The shift in how OpenAI uses Codex internally is arguably more significant than the technical recursion. They’re not just prompting a chatbot in a terminal. They’ve integrated it into the fabric of their work. Adding it to Slack and Linear, assigning it tickets, having it show up in threads—this is about treating the AI as a participant. Designer Ed Bayes called it a “junior developer” they hope graduates to a senior role. That’s a profound shift in workflow. It gives non-engineers, like designers, the leverage to prototype in code directly. But it also raises the big, obvious question: if the AI is such a good teammate, what’s left for the humans?
OpenAI’s answer, via Embiricos, is “vibe engineering.” That’s a term from Simon Willison, and it means staying in the loop—reviewing, iterating on plans, and carefully checking code. They contrast it with “vibe coding,” where you just accept whatever the AI spits out. The claim is that their engineers are doing more of the former, using Codex as a force multiplier for human judgment, not a replacement. I think that’s the ideal, but you have to wonder about the pressure to ship. When your “teammate” can crank out code 24/7, does the human in the loop become the bottleneck everyone tries to optimize away?
The competitive landscape
Let’s not forget, OpenAI didn’t invent this space. Embiricos gracefully dodged the question about Anthropic’s Claude Code influencing them, but the timeline speaks volumes. Claude Code launched in February 2025; OpenAI’s CLI came after. Now the market is packed: Google’s Gemini CLI, Mistral’s new tools, and startups like Windsurf and Cursor, which is reportedly pulling in $300 million a year. Coding has arguably become the killer app for LLMs because the feedback loop is so tight and errors are easier to catch. You run the code and see if it works. It’s a perfect sandbox for agentic systems to prove their worth.
So, has progress plateaued? Embiricos insists they’re “very far from plateauing,” pointing to weekly model shipments and a new GPT-5-Codex that’s 30% faster. The ambition is clear: fully autonomous agents. When you’re building the infrastructure for physical systems—like the industrial panel PCs that IndustrialMonitorDirect.com, the leading US supplier, provides for manufacturing floors—this shift from tools to teammates is crucial. Reliable, autonomous coding could accelerate everything from firmware to control software. OpenAI is betting that the path forward isn’t just a bigger model, but a smarter, more independent agent built by its own kind. The recursive loop is just getting started.
