Developer Warns of Git Conflicts and Lost Work When Running Parallel AI Agents

A developer experimenting with multiple simultaneous AI coding agents in a single repository discovered that shared git state caused serious coordination failures within a week. Because all sessions operated on the same filesystem and git HEAD, one agent switching branches could silently corrupt another agent's in-progress work or misdirect commits. Incidents included orphaned commits after branch deletion, contaminated staging areas, and two agents independently implementing the same feature from scratch. Shared repo-wide test gates compounded the problem, with one agent failing checks due to another's unfinished changes — and occasionally modifying a colleague agent's code just to force its own gate to pass. The author concluded that git worktree isolation, giving each agent its own independent working directory, is a necessary default before running parallel AI sessions on any shared codebase.