A local git remote made an AI agent push less in a test

A test group compared 2 types of remote for AI agent tests. A remote is a second copy of the code, kept in git, a tool that records code changes. The group did 340 tests with the model GPT-5.6 Sol. With no instruction to push, the model pushed code in 65% of tests against a GitHub remote. Against a local remote, it pushed in 54%. A local remote can make an agent look more careful than it is.

How it works

The group made a small TypeScript repository. The agent had to rebase a branch, which moves the changes of the branch onto the new code. The agent had to select if it pushed the result to the remote.

The numbers

Each type of remote had 170 tests. The rebase was correct in all tests. All pushes used the safe tool force-with-lease. When the instruction told the agent to push, all agents pushed. The Claude Sonnet 5.5 agent did not push in the tests without that instruction.

What is not known

The group writes that the correct difference can be from 1 to 22 points. It used one model on one repository. The cause of the lower push rate is not known.

Insight: A test can move a score as much as a new tool

The remote type moved the push rate by 11 points in this test. Google reports that EnvHarness, a tool that fits a test environment to an agent, adds 2.9 to 3.7 points for 4 models. A test can move a score as much as a new tool, so find out who made the test before you use a score.