I've been running multiple AI coding agents in parallel on the same repo. The agents code fine. The problem is managing multiple parallel sessions with git. Multiple agents, one branch, one overwrites the other.
Git worktrees give isolation but the per-session setup/teardown doesn't scale. GitButler looked promising but testing each agent's changes in isolation was tedious. Neither was built for this workflow.
git-stint automates the full lifecycle. Each agent session gets its own branch and worktree automatically. WIP auto-commits on exit or crash. Full diff, squash, and PR workflow for review. Conflicts caught before merge. Main stays clean.
Zero runtime deps. npm install -g git-stint. Built for Claude Code's hook system but works with any AI tool that writes to a git repo.
Been using it daily across my own projects. Would love feedback on the approach.
1 comment
Git-stint looks solid. One pattern worth adding: give each agent a dedicated worktree plus a lightweight session manager that persists terminal state across reconnects. Agents crash, networks drop, and losing mid-run context is brutal. If your orchestration layer can replay output and resume where it left off, you eliminate a whole class of "did it finish or not" debugging. Worktree isolation handles the git side; session persistence handles the human side.