Open Nine Terminals to Run AI Coding, and Use the Tenth to Check the First Nine

Someone opened nine terminals to run AI coding agents and eventually found that the only thing running in parallel was anxiety. This reply appeared under Emdash’s tweet. It did not get many likes, but it sums up the desktop state many people are in right now.

Emdash is an open-source desktop application under the Apache-2.0 license, with a very specific goal: let developers run multiple AI coding agents at the same time without switching back and forth between terminal windows. Each task has its own Git worktree and branch, whether the project is local or on a remote machine connected via SSH/SFTP. It is not tied to any one agent and directly invokes the CLI you already have installed, such as Claude Code, Codex, OpenCode, and Amp.

The features can be boiled down to a few points: isolated worktrees, so changes made by an agent stay on its own branch; issues and tickets from GitHub, Linear, and Jira can be sent directly into an agent; in-app diff viewing, CI checks, PR creation, and merging; and remote projects connected via SSH/SFTP. A commenter added that the flow in the README is to review the diff, open a PR, check CI, and merge in the app, with each agent’s changes going back to the main branch through the normal PR process.

In the comments, “worktree per task” is basically the consensus. Rohan Bhanot said this is the only sane way to run it, but merging back is still painful. Nash put it more concretely: four branches all touched the same configuration file, and the real work in the end was the merge. Another commenter mentioned that multiple agents crammed into the same checkout will fight over package-lock.json. Worktrees separate edits, but they do not separate merge order. Someone suggested marking task dependencies: if a dependency branch lands first, fail the whole batch; truly independent tasks can be merged in any order. That suggestion is very engineering-minded, and also very realistic.

When replying to another commenter, Dan Kornas said he wants the precise decision blocking a task to be exposed directly, so the question can be answered without reopening the entire session. This point is closer to the everyday pain than terminal management. When four agents are running at once, the time-consuming part is often not watching them write code, but figuring out which one has stopped and is waiting for you to make the call.

By gathering diff review, CI, PRs, and merge into one place, Emdash effectively puts merging back into the normal PR workflow. It will not automatically understand the semantic conflict when four branches change the same configuration file, nor does it promise to decide task dependencies for you. It solves the fragmentation of windows and context, not the wrap-up itself.

Another overlooked cost is status visibility. A commenter said terminals are buried under other windows, so completed tasks are often missed. Someone else asked whether agents keep running after an SSH disconnect, which the tweet did not elaborate on. At least Emdash centralizes tasks in a desktop app, making it more likely that you can see who is running, who has finished, and who is stuck.

The target user is clear: people who already pay for a coding agent subscription, have several fixes or features in flight at the same time, and do not want to manage terminal tabs anymore. Beginners working on a single task do not need to rush into it. The open-source repository is generalaction/emdash on GitHub. For now, do not open a tenth terminal; it may only be there to check the first nine.