Replace CrewAI with Parallel Claude Code Agents in Git Worktrees
Agent frameworks like CrewAI promise elegant multi-agent orchestration. Define your agents, assign roles, chain tasks, and watch them collaborate. In practice, the framework overhead often costs more time than it saves. There's a simpler approach - parallel Claude Code sessions running in git worktrees.
Why Frameworks Add Friction
CrewAI and similar tools introduce abstractions for agent communication, task delegation, and result aggregation. Each abstraction is a potential failure point. When an agent gets stuck, you debug the framework's orchestration layer instead of the actual problem. The indirection makes everything harder to reason about.
The Git Worktree Alternative
Git worktrees let you check out multiple branches of the same repo simultaneously without copying files. Each worktree gets its own working directory. Launch a Claude Code session in each one, give each agent a specific task, and let them work in parallel without merge conflicts.
Five agents working on five features in five worktrees. No framework. No message passing protocol. No YAML configuration files. Just terminals and git.
How to Set It Up
Create worktrees for each task with git worktree add. Open a Claude Code session in each directory. Give each agent a focused brief - one handles the API layer, another writes tests, a third refactors the data model. When they finish, merge the branches.
The key insight is that most "multi-agent coordination" problems disappear when you give agents isolated workspaces. They don't need to communicate if they're working on separate concerns. Git handles the integration.
When This Breaks Down
This approach works best when tasks are naturally parallelizable. If agents need to coordinate on shared state in real time, you need actual orchestration. But for most development work - feature branches, refactors, test suites - independent parallel agents outperform frameworks.
Fazm is an open source macOS AI agent. Open source on GitHub.