Git worktrees for AI coding agents

In late July 2026 developers highlighted practical benefits of using Git worktrees to run multiple AI‑driven coding agents. One report described how the jdupes utility could replace duplicate files with hardlinks, shrinking the disk footprint of several worktrees from roughly 9 GB to a much smaller size. It also introduced the open‑source Agent‑Up tool, which adds runtime isolation for each workspace’s processes, ports, Docker services and logs, helping to avoid conflicts when several agents share a monorepo.

A few days later an analysis pointed out that, despite these conveniences, Git worktrees do not constitute a true security boundary. Shared repository components such as refs, config, stash and hooks can be manipulated by an agent, allowing it to alter the main repository, rewrite commit identities and execute arbitrary code. The study noted that a fully isolated clone incurs similar costs to a worktree, challenging claims of near‑zero‑cost isolation. Concurrently, experts proposed a distributed repository network with regional mirrors to handle the high‑frequency operations of autonomous coding agents, aiming to preserve collaborative workflows while reducing latency and load on central Git hosting services.