Comparing Multiple Agents' Work with Git Worktrees & Cursor

In this video, I demonstrate a practical workflow for using multiple AI coding agents (Claude Sonnet 4.5 and Cursor's Composer 1) to build the same feature simultaneously, then comparing their work to choose the best implementation.

I'm building FilterHawk, a Gmail filter management app, and specifically focus on implementing the "action steps" feature (Phase 4 of the project).

Rather than relying on a single agent, I run two agents in parallel using git worktrees, allowing each to tackle the complex feature independently before reviewing both implementations.

The main focus is on my workflow for managing and comparing the completed work from multiple agents working in separate git worktrees. I walk through the challenges of reviewing agent work in worktrees—including issues with database credentials, migrations, and API integrations—and demonstrate a branch-based solution that makes the process more manageable.

You'll see me create separate branches (Actions-1 and Actions-2) for each agent's work, debug issues to get each implementation to a testable state, then evaluate both in the browser before merging the preferred solution. The video provides an honest look at the iterative debugging required even after agents complete their work, and shows how this multi-agent competitive approach helps ensure better results on critical features.

Topics covered in this video:

• Setting up multiple AI agents to work on the same feature using git worktrees
• The challenges of reviewing and testing agent work in separate worktree directories
• A branch-based workflow for isolating and comparing each agent's implementation
• Debugging and fixing issues to get each agent's work to a testable state
• Testing both implementations in the browser to evaluate quality and functionality
• Managing database migrations when switching between agent branches
• Merging the preferred implementation back into the main codebase
• When to use this multi-agent approach versus single-agent workflows