AI Pulse · Under the Radar

A developer named Cristi Cretu shipped a macOS app called diri last week that lets you run Claude, Codex, Cursor, and Gemini side by side, each working in its own isolated git worktree. The repo picked up 248 stars in seven days, which is notable for a tool with zero marketing budget.

What makes this interesting is the timing. Spotify opened public beta for Xirp in the same week, a system that does essentially the same thing at enterprise scale: swap AI agents mid-task across isolated environments without breaking state. Two independent teams, one solo and one inside a streaming giant, converged on the same core primitive within days of each other. That does not happen by accident. It suggests the pattern is becoming obvious to people building real workflows, not just writing whitepapers.

The practical piece is worktrees. Most people run one AI assistant in one repo and hope it does not wreck something three folders over. Diri and Xirp both isolate each agent in its own branch snapshot, so you can test four approaches to the same problem in parallel, pick the best one, and throw the rest away. No rebasing nightmares, no “undo, undo, undo” panic. You get clean rollback and clean comparison.

This is exactly the kind of orchestration layer we wire into the Omni Command Centre, where you need to run multiple models against live business context without risking your production data or losing track of which agent did what. The solo developer and the Spotify team both landed on worktrees because it is the simplest reliable way to keep agents from stepping on each other.

If you are still running one assistant in one window and manually copying results into a doc, you are working harder than you need to. The tooling to do this cleanly already exists, and it is moving fast.