Run Claude Code and Codex in separate Paneflow checkouts when their tasks can advance independently. Give each task its own file ownership, review its changes, then integrate and test both branches together. Splitting a terminal alone does not separate the files agents can edit.

## Start with two independent tasks

For a small first run, ask Claude Code to update `docs/setup.md` and Codex to add tests in `tests/parser.test.js` for an existing parser. Both can start from the same committed base. Agree on the parser's existing behavior first; if the tests require an implementation that does not exist yet, sequence those tasks instead.

Use Paneflow 0.17.4, a Git repository with a committed base, and installed, authenticated agents. Command syntax was checked with Claude Code 2.1.286 and Codex CLI 0.159.2 on Windows x64 on October 1, 2026. Authentication, model access and usage charges belong to each agent's provider.

## Choose the checkout before launching

1. Follow the [worktree reference](/docs/worktrees) to create `docs/setup-guide` and `test/parser-cases` from the same base, with **Worktree** enabled. Existing branches may reuse an existing checkout; inspect the path.
2. Open a terminal in each checkout. Run these commands in both and check that the directories and branches differ:

```sh
git rev-parse --show-toplevel
git branch --show-current
git status --short
```

3. Install the project's dependencies in the checkout that needs them. Use a distinct dev-server port per task, for example 3001 and 3002, configured through that project's own commands.
4. Run `claude` in the documentation checkout and `codex` in the test checkout. Enter the bounded task yourself. Do not ask either agent to modify the other task's files or integrate branches automatically.

The interactive entry points are documented in the [Claude Code CLI reference](https://code.claude.com/docs/en/cli-reference) and [Codex CLI reference](https://developers.openai.com/codex/cli/reference/). An unavailable command is an installation or PATH problem in that shell; check `claude --version` or `codex --version` there before retrying.

## Follow activity, then inspect the result

Paneflow's activity indicators depend on each agent's integration. An idle indicator is not proof that tests passed. Read the terminal, answer permission requests deliberately, and inspect the active checkout in **Changes**. It compares staged, unstaged and untracked changes against that checkout's `HEAD`; it is not a complete branch review.

Save editor buffers before review. Run the repository's relevant tests in the test checkout, inspect the documentation change separately, then preserve the reviewed work in commits. In your integration checkout, integrate those commits using your team's Git workflow, resolve conflicts and run the combined suite. Read the full branch diff as well as [Changes](/blog/diff-viewer). The observable result is two reviewed contributions and a tested combined tree, not merely two agents reporting completion.

## What was checked

This guide's shell-only rehearsal checks two Git worktrees, independent file changes, per-checkout inspection and integration on Windows x64. The installed agent versions and their help entry points were also checked. Paneflow's worktree creation/reuse was confirmed by the maintainer in the preceding content review; the [0.17.4 worktree implementation](https://github.com/arthjean/paneflow/blob/v0.17.4/src-app/src/workspace/worktree.rs) establishes the release behavior. No paid model task or speed comparison is claimed by this rehearsal.

## When to reduce parallelism

Two agents editing one file can overwrite or conflict with each other in a shared checkout. Worktrees keep working files separate, but commits can still conflict when integrated. Assign ownership or run the dependent task second.

Ports, databases, CPU, memory and provider limits remain shared. A port collision needs separate configuration, not another pane. If builds compete for memory, responses are rate-limited, or one task repeatedly needs the other's unfinished output, pause one task and sequence the work. Worktrees are not security sandboxes; local secrets and network permissions require their own controls.

Keep the [worktree reference](/docs/worktrees) open for checkout management and the [Files guide](/blog/files-sidebar) for save conflicts while reviewing the combined result.