Disclosure: This article is published by Hebbian AI, the team behind Dure, and was prepared with AI assistance using our public documentation. Product details were checked on September 14, 2026. The exercise below illustrates the documented workflow; it is not a benchmark.

Dure's public demo workspace. The two-task exercise below uses a smaller layout.

What problem does this workflow solve?

Starting a second agent is easy. Keeping track of which files it can change, what it has finished, and whether its work is ready to combine takes more care.

Dure brings agent terminals and code review into one workspace. A useful starting point is two small tasks with separate file scopes and explicit checks. Adding more agents can wait until that process works for you.

What is a Git worktree?

A Git worktree is a separate working directory attached to the same Git repository. It lets you work on another branch without switching the branch in your first directory.

For parallel coding, this gives each task its own checkout. It does not prevent merge conflicts when both tasks change related code, and it is not a security sandbox: credentials, processes, ports, and network access still belong to the same machine.

What do you need before starting?

The current Dure public beta supports Macs with Apple Silicon. Install Dure, then check that the agent command-line tools you want to use are installed and signed in. A command-line interface, or CLI, is a program you run from a terminal.

Dure does not include provider subscriptions or model usage. The public beta currently provides the Basic experience; Pro is not available. Follow the installation guide for current requirements and downloads.

How should you split the tasks?

Use a local Git project and choose tasks that can proceed independently. Here is an example exercise:

For example, one agent can receive this brief:

Review this project's installation guide. Check the documented commands, correct instructions you can verify, and list anything you could not check. Do not edit application code.

Assign either task to Claude Code or Codex according to your own preference. The separation comes from task scope and worktrees, rather than which model receives which task.

How do you start both agents in Dure?

1. Press ⌘N, choose the project, and describe the first task.
2. Enable Isolate in a dedicated worktree. Review the proposed folder and setup before starting the agent.
3. Repeat for the second task, choosing a different worktree and the other agent CLI.
4. Arrange the agent panes side by side in the same Space so you can follow both tasks.
5. Check each agent's working folder before allowing edits.

Splitting or moving a pane changes the layout; it does not create a worktree. Folder isolation is a separate choice when starting the task. The parallel-work guide explains this distinction and the setup steps.

How do you review and combine the results?

A diff shows the changes between versions of files. Open each task's Diff in Dure and check whether the changes stay within the requested scope.

Read the verification evidence as well as the code. An agent's completion message alone does not show that a behavior works. For the search fix, inspect the reproduction and test result. For the documentation task, check which commands actually ran and which steps remain unverified.

Send feedback to the relevant agent before combining its work. When both tasks are ready, combine the branches using your normal Git workflow, resolve overlaps, and run the checks needed for the combined result. Separate branches can each pass their own checks while still interacting badly after integration.

Frequently asked questions

Do Claude Code and Codex have to run together?

No. Start with one agent if the task is small or sequential. Parallel work is useful when the tasks can make progress independently.

Do separate worktrees guarantee conflict-free changes?

No. They separate working directories and branches. Changes may still conflict when combined, so keep the task boundaries clear and review the integrated result.

Does Dure pay for the agents' model usage?

No. You use your own installed agent CLIs and provider accounts. Check the relevant provider's subscription and usage terms separately.

The exercise follows Dure's public-beta documentation. You can find the product and its current demo at dureai.dev.

When you run parallel coding tasks, which part takes the most attention: defining the scope, following progress, or reviewing the combined changes?