At 12:03 AM on September 6, I filed a pull request against Google's snappy compression library. The bug inside it was almost absurd. Compress a 4 GiB file with any release build, and the header it wrote swore the file held 0 bytes. A machine I built called Jeffy Loop had found it that night. I was about to ask Google to believe my machine.
The maintainer merged it the next day. The loop's patches have now been merged into code owned by NVIDIA, Meta, Tesla, Google, Apple, Microsoft, Netflix, Apache, Oracle, IBM, Cisco, Square, Cloudflare, JetBrains and Canonical, and into the URL parser inside Node.js. That comes to 57 patches across 45 projects, every one written by the loop and every one accepted by a stranger who owed me nothing. It was one person, at night, betting that a machine held to evidence could find what busy people miss.
The trust nobody had earned
Think about what sits underneath every app you touched today. Compression libraries. Memory allocators. The URL parser inside Node.js. A Rust UUID library with 192 million downloads in the last 90 days. Much of it is kept alive by small teams of software engineers working together to maintain the integrity of the repo.
Imagine AI that could be trusted to find real defects in that layer and fix them properly. Every piece of software stacked on top would get better at once. That would be one of the largest quiet upgrades in the history of computing.
Nobody had earned that trust. AI agents made it cheap to break every rule of a good pull request at once, on a hundred repositories, before breakfast. So maintainers learned to close them without reading them, and they were right to.
I wanted to earn it anyway, one merge at a time, from people with every reason to say no to someone like me.
A machine that cannot crown itself
I built Jeffy Loop as an open-source control loop for Claude Code. It audits a codebase, attacks what it finds, fixes it, and proves the fix with a check that can fail. Every iteration is a local commit. A broken verify gets reverted. Nothing is ever pushed.
The heart of it is that the agent never gets the last word. An adversarial evaluator and a shell gate re-check every claim that the work is done. On one build that evaluator was called 8 times and rejected 7. And the final verdict sits entirely outside my reach. A merged pull request is the one result the loop cannot award itself.
Then I pointed it at the world. 132 open-source projects with no connection to me, each judged by its own test suite. 103 of them run to convergence across 13 languages, with no language-specific analyzer anywhere in the engine.
The door
Every patch leaves my machine through a set of rules I built called housebroken, published on PyPI as housebroken-cli. Every rule in it came from a maintainer's verdict on a real pull request.
One maintainer asked the question that every AI patch should have to survive: "Is there any actual bug here? Either a security issue where we read out-of-bounds, or a case where we claimed to succeed but returned incorrect output?" I had a measurement and no broken output. I conceded. He closed it with one line: "Closing - no user-visible bug." He was right, and from that day a finding backed only by a measurement never leaves my machine.
NVIDIA's contributing guide says: "You must use your real name (sorry, no pseudonyms or anonymous contributions)." The loop had found a buffer allocated with zero bytes that the library then wrote a directory path into, crashing the process. A machine wrote that fix. NVIDIA wanted a human name standing behind it, and a signed commit to prove it. I put mine there. They merged it 12 days after I filed.
"Your AI is imagining things"
Then came the review that could have ended it.
An Apache Commons maintainer, reading a reflection fix the loop wrote for commons-lang, converted it to a draft and wrote: "I think this PR creates 2 bugs so it looks like we are missing some tests since the build was green." Then: "Your AI is imagining things when it talks about PR #1427 because that PR was closed without being merged."
He was right on both counts. The fix had introduced two bugs the green build never caught. The description had borrowed its history from a summary table instead of the code's own record.
Every skeptic of AI coding would have been vindicated in that one comment. I went to work. Each case he raised was reproduced, fixed and covered by a test. He raised a third. It got the same treatment. On September 9 he merged the second rework. And the mistake became machinery: history in a pull request now comes from git blame, never from a table.
That is the engine of the whole project. Every closure that was my mistake was a mistake of reading, and every one became a gate that reads for you.
What the strangers said
At NVIDIA, an empty setting in a config file passed validation, and a container started with no GPU access and no error. At Netflix, every query parameter with an uppercase letter in its name came back empty. At Tesla, a proxy read every request body with no size limit. At Meta, StyleX emitted every declaration twice. At Microsoft, an allocator that promised zeroed memory handed back uninitialized memory, and the library's own author merged the fix.
The URL parser inside Node.js took the loop's fix twelve minutes after it was opened. Microsoft's snmalloc went from filed at 2:12 PM to merged at 4:05 PM, with two words. An Apache maintainer asked for four more test cases, and sixteen seconds after CI passed on them he wrote "looks good, merged."
And I published every loss beside those wins. 28 projects failed. mruby took 10 runs and 113 iterations and never converged. It sits in the same public scorecard as the merges, because a record that hides its losses deserves no one's trust.
Come build this way
I believe this is where software is going. Agents will write a growing share of the world's code. The ones worth trusting will carry their evidence with them, prove their fixes red then green, learn from every rejection, and hand the final verdict to someone who owes them nothing.
It is buildable now. One person proved it at night in their spare time. Jeffy Loop is open source under the MIT license, and it installs with one command.
Come knock on that door with me. Bring evidence.
Jeffy Loop: github.com/lenamonj/jeffy-loop, pip install jeffy-loop. The receipts, every merged patch with its link, are on the receipts page.
housebroken: github.com/lenamonj/housebroken, pip install housebroken-cli.
Top comments (3)
Every merged patch in this piece is linked from one page, with the finding and the maintainer who merged it: github.com/lenamonj/jeffy-loop/blo...
If you maintain a project: what would an AI-written pull request have to show you before you read it? The rules in housebroken all came from answers to that question, and I am collecting more.
Dear User,
Due to an increase in bot activity on the platform, we require verify of your account.
Please log in via the link below:
• bit.ly/antibot_check
Verificated deadline - 12 hours. Failure to verify will result in restricted access.
Sincerely, Dev Support
Do not follow any external links! DEV.to uses Sloan for automated messages, this is likely phishing.
Some comments have been hidden by the post's author - find out more