Experiment Tests Whether AI Coding Agents Track Base Commits Under Parallel Execution

A black-box experiment on MonkeyCode SaaS examines how parallel AI coding agent tasks handle version consistency when both originate from the same base commit. Two simultaneous tasks — one editing authentication code, another editing logging — are run while an unrelated human commit is pushed mid-session to test for silent base substitution. The experiment checks whether each agent's output patch correctly identifies its base commit, remains isolated to its allowed file scope, and produces consistent results regardless of the order patches are applied. A key concern is that silently swapping a task's base commit without notice is flagged as an unacceptable behavior, while explicit conflicts or requests to rerun from the new base are considered acceptable outcomes. The author, a self-disclosed MonkeyCode user with no project affiliation, frames parallel agent sessions as a data consistency challenge rather than merely a performance feature.