How to Effectively Run Many Claude Code Sessions in Parallel

How to Effectively Run Many Claude Code Sessions in Parallel

Running coding agents sequentially rather than in parallel means you're missing out. One of the key benefits of coding agents is the ability to complete work in parallel, something that was not feasible before in software engineering tasks.

However, when I run a large number of parallel coding sessions, it's often difficult to keep track of them. You need specialized techniques to easily manage an overview of all the sessions, quickly catch up on conversations when returning to them, and more.

This infographic highlights the main contents of this article. I discuss how to effectively run many coding agents in parallel, addressing why it's a challenging task and the techniques I use. Image by ChatGPT.

Why Running Coding Agents in Parallel is Challenging

First, I'll cover why running coding agents in parallel is difficult. To some extent, this should be self-evident. Before LLMs, software engineers and programmers would naturally work on one task at a time. The reason was that attempting multitasking on software engineering tasks would reduce effectiveness. Software engineering tasks are often complex and require full concentration. If you try to do other things simultaneously, performance on all tasks is likely to suffer.

However, the landscape has changed significantly. You no longer write all the code yourself. In my opinion, you shouldn't write all the code yourself; coding agents should handle that. As a programmer, you should now manage coding agents rather than writing code directly.

As a manager of coding agents, you naturally have to handle agents performing different tasks. Performing tasks in parallel inherently requires these tasks to be independent. This presents the same challenge as before, where you need to manage multiple programming tasks simultaneously, keeping a lot of context in your working memory. You need to be able to:

• Respond to questions from your coding agent.
• Test the coding agent's implementation after it's done. You need to know what to test, how to test it, and how to ensure the coding agent performed as expected.

How to Effectively Run Many Parallel Coding Agents

In this section, I'll cover some specific techniques I use daily to effectively manage many parallel coding agents.

Agents View in Claude Code

This image shows the agent view in Claude Code. It's a simple terminal view where each task you're working on appears as a single line, with Claude clearly marking which tasks are running in the background and which require your input. This makes it much easier to manage many agents compared to having one terminal tab or pane per agent. Image by the author.

A more effective technique for managing an overview of your agents is using an agent view, such as the one available in Claude Code. Different providers have various ways of presenting this. I know Warp, the terminal, has also recently introduced a new method for viewing multiple agents.

The advantage of the agent view is that, as shown in the image above, you don't need to full-screen each conversation with an agent. Each conversation becomes a single line, which you can expand by pressing enter if you need more details. Otherwise, it remains a background task and will prompt you for input whenever any Claude session requires it.

I find this an effective way to work, as you can easily manage many different coding sessions without confusion about which agent is running or which needs input. You can activate the agent view in Claude Code with:

claude agents

Alerts When Coding Agents Need Input

This image shows my Warp terminal when Claude Code instances need input from me. You can see the star between the Claude Code symbol and the title of each tab, indicating that the session requires input from me. If there's no star, the process is running in the background, and I don't need to interact with it. Image by the author.

Another important technique is being alerted whenever coding agents need input. If you use agent mode, this is clearly indicated whenever a coding agent requires input. However, if you prefer not to use the agent view or want an alternative, there are different options.

In the image above, you can see different Claude Code sessions running in separate terminal tabs. A star appears between the tab title and the Claude Code icon if the session needs input. This is a simple way for me to identify which tabs require action and which are running in the background.

• * *

You can also set up an audio signal to alert you whenever an agent requires input from you. For example, you can implement this using hooks in Claude Code, which are processes that execute at specific times. A hook can be triggered each time Claude needs input from you, and you can link this hook to an audio signal, so you are notified whenever one of your coding agents needs input.

Activate recaps

This image illustrates the recap feature in Claude Code. A recap is a summary of your activities and goals in a particular thread. This helps you quickly regain context, which is especially useful when working with multiple agents simultaneously. Image by ChatGPT.

Recaps are another powerful feature that enables you to manage numerous parallel coding agents effectively. A common challenge when running multiple agents is keeping track of the context for each one.

Suppose you have five agents running concurrently. You initially interact with agent 1, provide instructions, and then launch agents 2, 3, 4, and 5. By the time you return to the first agent, it may have been 10 minutes or longer since your last interaction. You need to quickly recall what you were doing with that first agent and what your objectives were. Without a recap or similar tool, this can be challenging, but the recap feature in Claude Code is very helpful.

As shown in the image above, the recap is a text snippet located above the user input field. It summarizes your actions and goals in that thread. By reading this text, you can quickly regain context and continue interacting with your agent.

New tabs or split panes

This image shows a split pane with multiple instances of Claude Code running. Split panes are extremely useful because they allow you to monitor several coding agents at once without switching between tabs. My typical setup involves one tab per repository, with split panes within each tab to provide an overview of all agents. Image by ChatGPT.

The final technique I want to cover is the use of new tabs or split panes. I strongly recommend using a terminal or another platform that supports split panes when working with coding agents.

The image above demonstrates a split pane setup. In Warp terminal, pressing Command+D on a Mac splits the current view horizontally, giving you two separate terminals. This is very useful as it allows you to quickly review two agents at once, which I find beneficial.

Many terminal providers and coding platforms offer split pane functionality, so I suggest finding one that suits your workflow.

• * *

My preferred method is to open a new tab for each folder I work in, so I have one tab per folder. If I am working with multiple agents within a folder, I use split panes.

This setup allows me to quickly review my coding agents across different repositories.

Conclusion

In this article, I discussed strategies for maintaining an overview of parallel coding sessions. I covered various techniques for efficiently running multiple agents while keeping track of all of them. I believe the future of programming involves orchestrating coding agents. You will become a manager of AI agents. You should start developing your skills in managing multiple coding agents in parallel, as I believe this will be a crucial skill for programmers in the future.

Get in Touch

 My free eBook and Webinar:

10x Your Engineering with LLMs (Free 3-Day Email Course)

Get my free Vision Language Models ebook

My webinar on Vision Language Models

 Find me on socials:

Substack

LinkedIn

X / Twitter