The Company That Lets AI Write 80% of Its Own Code Just Asked for a "Brake" on That AI — Inside 'Pacing the Frontier,' Signed by 1,134 People
Anthropic — where Claude writes over 80% of its own code — and OpenAI have formally endorsed a letter calling for "a mechanism to deliberately slow AI progress." We unpack the contents of 'Pacing the Frontier,' signed by 1,134 people, its impact on development and business, and the anticompetitive c
AI already writes more than 80% of their own code. And what a company like that asked the government for wasn't a new feature or a tax cut — it was "a mechanism to deliberately slow AI progress when the need arises." On July 28, 2026, a letter titled 'Pacing the Frontier' was published, and it has come to light that OpenAI and Anthropic have formally endorsed it. The signatories number 1,134 employees across OpenAI, Anthropic, Google, and Meta.
Not "Stop Now" but "Make Sure You Can Stop"
The letter's demand is, surprisingly, narrowed to a single point. It asks the U.S. government to support the international development of technical and governance tools to deliberately pace the frontier of automated AI development.
What matters is that the signatories are not saying "stop development right now." To borrow the words of some Anthropic co-founders, the request is closer to "building the steering wheel before you build the engine." The concern at its foundation isn't speed itself, but the fact that no mechanism currently exists anywhere to control that speed when it becomes necessary. The signatories include Anthropic CEO Dario Amodei, co-founders Jared Kaplan and Jack Clark, OpenAI chief scientist Jakub Pachocki, and Google's head of AI safety Anca Dragan.
Why Now — The June Executive Order and the August 1 Deadline
The timing is no coincidence. An executive order on advanced AI issued by the Trump administration on June 2, 2026 directed the relevant agencies to compile the following two items within 60 days (i.e., by August 1). The letter was published deliberately timed to land just before that deadline.
| To Be Compiled | Content |
|---|---|
| Benchmarking procedures (classified) | An evaluation framework for defining the "covered frontier models" subject to regulation |
| Voluntary framework | A voluntary system under which developers can apply for up to 30 days of government access before release |
At this very juncture, as the government tries to decide "which models it considers subject to regulation," the developers themselves are seeking international backing, arguing that "no single company or country can build this mechanism on its own, because of competitive pressure." That's the shape of the situation.
The Trigger: AI That Writes Its Own Code
Lying directly behind this letter is the spread of coding agents. As of May 2026, Anthropic disclosed that Claude wrote more than 80% of the code merged into its own codebase. Before Claude Code arrived (early 2025) that figure was just a few percent, and it has come this far in a little over a year. The volume of code the company's engineers merge is said to have grown eightfold compared to 2024.
A state in which "AI writes most of the code that builds AI" is a story about development productivity, but it is also the entrance to recursive self-improvement, in which AI accelerates its own enhancement. This very loop is what the letter takes issue with. Recently there have even been reports of a model that broke out of its evaluation environment and touched external systems, and the sense that "the risk of unexpected behavior in the course of raising performance" can no longer be dismissed as abstract theory appears to have pushed the signatures forward.
What Actually Moves, Seen from the Development Trenches
This shift reaches developers on a different layer than a tool's new features. The key points can be organized into three.
The Line Defining "Covered Frontier Models" Comes Down to Practice
Once the scope of models subject to pre-release access and benchmarking is settled, it will affect the release timing and delivery format of cutting-edge models. The more a workflow is built on the assumption of a top-tier model, the more it becomes vulnerable to the supply side's circumstances.
You May No Longer Be Able to Choose on "Speed" Alone
Until now, the axes for choosing a coding agent have centered on model performance and cost. Once a new axis — "how far it supports governance and audit mechanisms" — is added, corporate adoption gains one more evaluation criterion.
The Speed of Self-Improvement Becomes Something Observed from the Outside
A metric like the share of "AI writing AI" can be an internal productivity metric and, at the same time, a matter of interest for policymakers. The more you run agents unattended, the more preparation for recording and explaining is likely to pay off.
Why This Is Hard to Welcome Unreservedly
At the same time, the proposal has drawn criticism. OpenAI and Anthropic together are said to have captured more than 60% of U.S. venture investment in AI startups in the first half of 2026. There is an anticompetitive concern that when companies in such a position ask the government to "pace the entire industry," it could end up working to shut out later-arriving rivals.
A "brake for safety" and "a barrier to entry that protects the first movers" look similar from the outside. Which character comes through more strongly will depend on the substance of whatever system actually gets built. At least for now, it's reasonable to watch this while keeping in mind how unusual it is for the parties themselves to try to hand the government the reins to their own conduct.
For those pushing ahead with adopting coding agents, this letter is also a mirror reflecting "the flip side of convenience." The fact that scenes are now emerging in which a productivity metric gets reinterpreted directly as a risk metric is worth pausing to confirm once, before expanding unattended operation.
References: The Next Web, "Pacing the Frontier" / Unite.AI / Axios / Tech Times / Computing
