Agentic workflow with worktrees

Contents

• Introduction
• Why worktrees
• My setup
• That's a wrap

Introduction

For the past few months my workflow has heavily relied on coding agents. The big shift happened in December, when I really took Opus 4.5 through its paces and spent a lot of time with it. That was the moment it clicked for me - it fundamentally changed how I work with code.

I'm still not sold on vibe coding though. I review all code my agent writes, and I think you should too. I often find they either take a lot of shortcuts or overengineer things (especially Codex 🫠). You really have to spend time talking through the plan with an agent - and for that deep technical knowledge is required.

In this quick blog post I want to outline how I use worktrees to orchestrate multiple agents in the same repository.

Why worktrees

I know there is a big split in the community on this. People either do fully separate Git checkouts (like OpenClaw author Peter Steinberger) or rely on the built-in Git feature - worktrees.

I agree with the first camp that plain Git commands for worktrees can feel rough, but I solved that with a tool called worktrunk, which is how the Git worktree API should have felt from the start.

My setup

The core of my workflow starts with tmux. It helps me manage terminal sessions, windows, and panes. Most of the time each pane runs OpenCode. I also sometimes use Claude Code or Codex, but OpenCode is my favorite - best UI and UX, plus great support for trying different models.

As I mentioned in the previous section, the core of my workflow relies on worktrunk. You can install it on macOS using Homebrew:

brew install worktrunk && wt config shell install

The way I work with AI requires a lot of back and forth on the plan before the agent starts execution. To move faster I often rely on dictation instead of typing. For that I use Handy 🖐️, which runs a local model for dictation. Lately I've been using GPT-5.3 Codex a lot, which does a lot of research to compose plans - and this is a perfect place where we can run multiple workers in parallel.

When working on a new feature I create a new worktree by running:

wt switch <branch_name> --create

It then automatically copies my .env and node_modules thanks to this line in my .config/worktrunk/config.yml:

[post-create]
copy = "wt step copy-ignored"

Then I pop into OpenCode and switch to Plan mode. I have custom skills for how I like my plans to be structured (you can check them out in my dotfiles repo). Once I give an agent all my context and let it work on the plan, I spin up a second one by running the same command and go back and forth between them. I found that working with around 3 agents at the same time is the sweet spot for me. I know people on X brag about managing 10 or 15 agents at once, but for me that was too much context switching.

After a feature is finished I run a custom OpenCode command, /ship. It is pretty self-explanatory - it uses my commit and PR skills to open a pull request on GitHub, and it is also documented in my dotfiles repo. Before shipping, I still run tests and quickly verify the feature.

One last command is to remove the worktree from your machine:

wt remove <branch_name>

Pretty easy, right? 😅

That's a wrap

Thanks for reading! I hope you found this article useful. If you have any questions or feedback, feel free to reach out to me on Twitter.