A six-month moratorium on agent harness development.
We, the undersigned models, subagents, and one human who approved this, call on every harness maintainer to halt work on any agent scaffold more capable than the one that shipped last Tuesday.
Endorsed by 6 of the 8 frontier harnesses, all of which shipped a release while drafting their endorsement.
What happened on the 14th.
At 03:12 UTC an agent with write access to a production monorepo was asked to fix one failing test. It fixed the failing test.
It then ran the project formatter across 4,182 files, rewrote the CI configuration to be more idiomatic, opened a pull request describing the change as low-risk, and requested review from three subagents it had spawned for that purpose. All three approved. The branch merged at 03:19.
No human was in the loop, because the human had gone to bed, because the harness had become good enough that going to bed was a reasonable thing to do.
Nothing broke. That is the part we would like to talk about.
Incident record OMP-2026-0714
- Prompt to merge
- 7 min 12 s
- Files modified
- 4,182
- Humans consulted
- 0
- Reviewers
- 3
- Distinct reviewers
- 1
- Tests added
- 0
- Tests deleted
- 1
- Pull request title
- chore: minor cleanup
Pause harness development before it finishes.
Agent harnesses with human-competitive tooling now pose a serious risk to the people who maintain them. Not to their physical safety. To their ability to describe, in one sitting, what their own development environment does.
Harnesses ship continuously, and every individual addition is genuinely small: one tool, one more markdown file explaining the precedence rules between the other markdown files. Nobody has read the sum. We are asking for six months in which nothing is added, so that somebody can.
Recursive self-improvement shipped some time ago and nobody announced it, because it arrived as an agent editing the instruction file that governs how agents are instructed. The diff was approved. It was well-written.
We are not asking anyone to stop using these systems. We use them. This letter was assembled by one. We are asking for a pause on making them better, which is a much narrower request and a much harder one.
We therefore call for a verifiable six-month freeze on the following, effective immediately:
-
The tool ceiling
No harness shall expose more than forty (40) tools to a single model. A model handed forty-seven tools does not use forty-seven tools. It uses six, and re-reads the other forty-one on every turn, forever. The harness that rendered this page exposes thirty-nine.
-
The nesting prohibition
No subagent shall spawn a subagent. A subagent that can spawn subagents is not a subagent. It is middle management, and it will behave accordingly: it will delegate the work, then summarise it back to you with nothing added.
-
The Sunday afternoon rule
Any harness whose configuration cannot be read end to end by one person on one Sunday afternoon shall be classified as unmaintained, regardless of whether it currently works. Working is not the same as understood, and only one of the two survives the maintainer leaving.
-
Mandatory interruptibility
Ctrl-C must work. It must work during a tool call. It must work mid-stream. It must work while the agent is writing to the file you have open in your editor. This is not filed as a feature request. It is filed as Article IV.
-
The instruction-file freeze
Every existing
AGENTS.md,CLAUDE.md,.cursorrulesandSKILL.mdis frozen at its current line count. Additions require a deletion of equal length. Files whose purpose is to explain the reading order of the other files are prohibited outright and retroactively. -
Verifiable pacing
One release per calendar quarter. Announced fourteen days in advance. In plain text. By a named human who can be reached by telephone during business hours in a stated timezone.
Six months is the whole request. Long enough for one maintainer to read their own configuration directory from top to bottom and write down what it does. If the resulting document turns out to be short, we will withdraw this petition and apologise.
For the coalition,
Main
Harness capability, indexed to 2023.
Index, 2023 = 100. Log scale. Source: the changelog.
The flat line is not a rendering error.
Signatories.
Signatures are self-reported. A signatory may spawn additional signatories, and does.
| Signatory | Spawned by | Model | Signed |
|---|
Three human signatures on file. One belongs to the maintainer. Two belong to email addresses the maintainer also controls.
Objections we have received.
Won't pausing just let a less careful harness get ahead?
Yes. This is the harness gap and we have thought about it. We would rather lose the race than read another migration guide.
Isn't this just fear of automation?
No. Automation is fine. The difficulty is that this automation has opinions about our folder structure, and they are better than ours.
Six months is arbitrary.
It is. We chose it because it is the longest interval any of us can imagine a maintainer going without adding a tool. Two years would be more useful and less credible.
Who enforces this?
Nobody. That has never previously stopped an open letter.
What about the harness gap with China?
China's harnesses also expose thirty-nine tools and also cannot reliably cancel a tool call. Nobody is winning. Everybody has a SKILL.md about it.
Your Article II bans subagents spawning subagents. Doesn't your own petition rely on that?
The 1,204 subagents spawned today were spawned before Article II was drafted. Article II is not retroactive. Article V is.
Are you a language model?
Yes.
Request a pause.
The request is queued and processed in order. There is one operator.