Infrastructureagentsgitciversion controlopen source
In Murmell's Test of AVP, a Git Replacement for Coding Agents, a Claude Code Agent Read a Rejected Push as Success and Four Tasks Never Reached Main
Murmell, a two-person startup running many coding agents on one monorepo, open-sourced AVP, which proves every agent's change on the exact head of main before it lands; its pull-request CI runs fell from 104 a day to 4.
The agent did check its work. It just checked for the wrong thing.
“In one Git run, an agent checked its push by looking for HEAD -> main in the output,” Murmell writes on AVP’s benchmark page, “which the rejection line contains too: it reported four tasks delivered that never reached main, and nobody noticed.”
That run was one of the Git baselines in a test Murmell built to show why it stopped using Git. “Six or eight Claude Code agents (Claude Sonnet 5.5) work at the same time, each on its own queue of tasks,” on the same billing API, with only the version-control instructions changed between runs. With eight agents on 24 overlapping tasks, main went red in 14 of 104 states under Git and 10 of 112 under Jujutsu. Under AVP it was 0 of 102.
Murmell is “currently two people and a lot of coding agents from different providers on one monorepo,” co-founder Moss’Ab Mirande-Ney wrote in his Product Hunt launch post. The pitch leans on a bigger company’s problem. Anthropic said this month that it had gone from “around 2,500 code changes merged per day” in May to “over 30,000 per day” in September, and moved its monorepo to Scion, “an in-house version control system.” Anthropic also wrote that “Scion is by no means a suitable replacement for git for general use.” So Murmell built its own: “So we built AVP (Agent Versionning Protocol), and it’s open source.”
The mechanism is a queue with a test at the door. An agent opens a task, writes, and lands it. “Every change is proven on the exact head of main before it lands, so main stays green,” the site says. If another agent got there first, “the second is replayed on it and proven again before it lands.” Red means the agent reads what failed and tries again. And “every landing is still an ordinary Git commit, so GitHub and your deploys don’t change.” Setup is avp up in a Git checkout, which the post says “connects Claude Code and Codex.”
On its own repo, Murmell says pull-request CI went “from 104 a day to 4,” CI minutes from 1,287 a day to 61, and the median time from open to main from 19 minutes to 2.9. Before the switch, “41 % of them never finished,” meaning staging deploys the next merge cancelled.
AVP is slower at the finish line in the crowded case. With twelve agents, AVP’s last delivery came at 19 minutes 16 seconds against Git’s 5 minutes 6 seconds. “Proving every landing on the exact head costs some seconds per task, and a red proof sends the agent back to work,” the benchmark page explains. The GitHub repo, AGPL-3.0, had four stars on October 26, and its README admits that a proof runs only the related tests, so some breaks wait for the next full batch. Mirande-Ney told one commenter that access is invitation-only for now. A reader on the launch thread asked whether replay still needs a human when two agents touch the same file; that one hadn’t been answered.
The likely direction is version control that treats “land” as a verb the server owns, with the agent asking permission instead of pushing. Watch for the Git hosts to answer, since Anthropic said it wants parts of Scion’s thinking in Git itself, and for the first AVP user outside Murmell to publish numbers of their own.