Parallel Coding Agents with Container Use and Git Worktree
You’re going to love Container Use.
I’ll explain what it is and how to use it for development.
Trust me this is a good one. The Container Use demo starts about half-way through.
Check out the video on YouTube
In the demo I use Claude Code, but this tool (and the two other strategies that I’ll show you) work with pretty much any coding agent (cursor, gemini, etc…)
Get started with container use here
https://container-use.com/introductionTimestamps:
1:10 - Copy Folders Guide
5:23 - Git Worktree Guide
11:55 - Container Use Guide
Topics from this week’s video
Parallel agents for generative coding
Three workflows demoed with Claude Code (also applicable to Cursor, Gemini, etc…)
Goal: safely build multiple feature variants in one codebase
Workflow #1 — Copy folders
Duplicate the project into Agent A/B, run the same task, compare results
Pros: quickest to start
Cons: dependency headaches, heavy disk use, easy to trample main, manual cleanup
Outcome: pick the better variant, discard the other, sync with main
Workflow #2 — Git Worktree
Separate working directories tied to their own branches for each agent
Build and compare variants (e.g., CTR-focused vs. aesthetics-focused), merge the winner to main, remove extras
Pros: proper branching, lighter footprint, clearer history
Cons: each worktree still needs its own environment setup
Workflow #3 — Container Use (Dagger)
Claude integrates via MCP; each task runs in an isolated, disposable environment with automatic commits/branches
Local code stays untouched until you pull changes; simple diff/merge/cleanup; plays nicely with CI/CD
Demo: multiple parallel redesigns (RPG, Barbie, Minimal); one broken variant was auto-fixed in a fork and chosen; merged and deployed
Pros: strongest isolation, reproducibility, true parallelism
And the end of all our exploring
Will be to arrive where we started
And know the place for the first time.T. S. Eliot, Little Gidding
Container Use is brilliant for the isolation piece. The three workflows you've laid out cover the external tooling side really well.
Since you published this, Codex has added a native multi-agent mode that handles some of the same problems from the inside. Instead of managing containers or worktrees yourself, Codex spawns sub-agents with role-based configs. Your explorer runs on a fast model in read-only mode, your worker gets full write access on the heavier model. Each gets an isolated context automatically. I covered the setup here https://reading.sh/codex-has-a-multi-agent-mode-and-almost-nobody-is-using-it-088e44f774ef if you're curious.
Would be interested to see how Container Use + built-in multi-agent might stack. The container isolation for the environment plus the context isolation from sub-agents could be a solid combo.