OpenAI Hired the 'Git AI' Founders to Prove Codex's ROI — the Coding-Agent War Just Shifted From Capability to Receipts

For two years the coding-agent race ran on a single number: how much code your agent could generate. Lines per day, pull requests per sprint, "vibe" shipped. That contest is now officially over. OpenAI just hired the two founders of Git AI, an open-source tool that tracks how much code agents actually write and, more to the point, what happens to that code afterwards. The message is unmissable: the next battleground is not capability, it is receipts.

I have written before about the trust problem in vibe coding and the deployment crisis nobody likes to discuss. This hire is the other side of the same coin. When the vendor selling you the coding agent starts buying the measurement tools, pay attention to what they think you are about to ask next.

What Git AI actually did

Git AI is one of those projects that began as a side question and turned out to be the question everyone needed answered. Founders Aidan Cunniffe and Sasha Varlamov started building it in the summer of 2025 because they wanted to know something simple: how much of their own code was being written by AI, and what happened to that code once it was committed.

The answer was not obvious. So they built a Git extension that links every line of code back to the agent, model, and prompt that produced it, then tracks that line as it moves through the pipeline. How much reaches production. How much gets rewritten. Where the tokens go to waste.

It supports the usual suspects: OpenAI Codex, Anthropic's Claude Code, Cursor, Google's Gemini CLI, plus the background agents such as Devin and Codex Cloud that keep churning while you sleep. The pitch to developers was beautifully simple: just prompt and commit. Keep your existing tools. Git AI quietly records the split between human and machine code in the background.

That is the thing OpenAI now owns, or at least the two people who built it.

Why OpenAI made this move

The official line is that Cunniffe and Varlamov will join the Codex team and help businesses understand Codex's impact. OpenAI's Thibault Sottiaux was direct: the company wants to use Git AI's technology to help organisations see where Codex is making a difference.

Read between the lines and the strategy is plain. OpenAI is not buying a feature, it is buying an argument. The argument is that Codex does not merely write more code, it writes code that ships, and here is the proof.

There is a reason this matters right now. The enterprise market, where the real money lives, stopped asking "can your agent code" about a year ago. Every agent can code. The new questions are harder. What does this cost us per merged pull request? Which agent is actually cheaper once you factor in rework? How much of what it generated is still in our codebase three months later?

Those are ROI questions, and until now nobody could answer them with anything better than a gut feeling and a marketing deck. Git AI was built to answer precisely those questions, for every agent, not just Codex. That last part matters a great deal.

The receipts era has arrived

Here is what I find most interesting, and it goes well beyond OpenAI. The entire vibe-coding movement has been running on faith. Faith that the agent is saving time. Faith that the code is decent. Faith that the tokens you burned translated into value.

Faith is fine for a side project. It is not fine for a chief financial officer signing a six-figure seat contract.

I think we will look back on this hire the way we look back at the first time a SaaS vendor published a public uptime page. It seems like a small thing. It is actually the moment the product got honest about what it does and what it does not.

What Git AI represents is the arrival of a measurement layer for AI-assisted coding. Once that layer exists, and once a major vendor has folded it into their flagship product, the dynamic flips. Agents stop competing on how much they can generate and start competing on how little they waste. Rework rates, production reach, token efficiency. Those become the spec sheet.

I suspect a number of the more breathless AI-coding tools are about to have an uncomfortable quarter. If your agent can generate a thousand lines a minute but half of them get rewritten by a human before merge, the new tooling is going to make that visible, in public, to the person holding the budget.

What this means for you, the builder

You do not need to care about OpenAI's internal politics. What you should care about is the lesson, and it transfers cleanly to the no-code and agent world.

You are going to be pitched a lot of AI agents in the next year. Some will automate your support queue. Some will write your content. Some will build your internal tools. Every vendor will show you a demo of the agent doing something clever. Almost none will show you the number that matters: what did this actually save or produce, and can I prove it.

One reason this transfer is so clean is that no-code builders have been living with a version of this problem for a decade. Every low-code platform promises faster delivery, but the teams that last are the ones that measure time-to-live, not time-to-demo. An agent that drafts a support reply in four seconds means nothing if a human still has to read and rewrite it every time. An agent that drafts and a human approves in one tap is different economics entirely.

So here is my advice, and I mean it. Treat ROI proof as a requirement, not a nicety. Before you adopt any agent, or any no-code platform claiming to ship agents, ask for the same thing OpenAI just decided to build: a way to see where the agent made a difference versus where it burned your budget.

Concretely, that means a few specific asks. Can I see which outputs reached production or were actually used? Can I see rework, the places a human had to step in and fix what the agent did? Can I see token or run costs per completed task, not just per month? If the vendor cannot answer those questions, what they are selling you is faith, and faith is not a line item.

The takeaway

The coding-agent war has moved from "who writes the most code" to "who can prove the code pays off," and OpenAI just telegraphed that shift by hiring the two people who built the measuring stick. The same shift is coming for every agent and no-code platform you touch.

Do not wait for your vendors to volunteer the numbers. Demand them now, while you still hold the leverage, because the vendors are already building the tools to make their case to the next person who asks. Make sure you are the one asking, not the one being sold to.