Common information highlights how multitasking needs to be prevented as a lot as attainable as a result of it removes you from focus. Nevertheless, with the rise of coding brokers, parallel work has turn out to be a requirement to be environment friendly. Contemplating coding brokers can run for an extended time period, you don’t need to spend time ready on them to complete. As an alternative, you need to fireplace off an agent, and whereas that agent is working, you need to begin one other process.
Nevertheless, beginning off duties in parallel isn’t easy. There are quite a few concerns it’s essential to make, comparable to
- How do you run a number of coding brokers in the identical repository?
- How do you continue to reduce context switching as a lot as attainable?
- How do you retain an summary of your entire brokers which might be operating?
I’ll cowl all of those factors in at the moment’s article with specifics on how I take care of these issues and the answer I apply.
Why run brokers in parallel
As with quite a lot of the opposite articles I’ve written, the primary purpose for the methods I describe within the article is saving time and being extra environment friendly as an engineer. If you wish to get probably the most out of coding brokers and the unimaginable LLM improvement we’ve seen in the previous couple of years, it’s essential to parallelize your work.
To grasp why, let’s think about the alternative instance. Let’s think about solely working sequentially, such as you usually did when programming earlier than LLMs turned obtainable. Sequential programming with coding brokers would look one thing like:
- Discover the duty you need to do
- Describe it to an agent and make a plan
- Begin the agent and let it run. Wait till it finishes or asks you for one thing
- Check implementation and iterate
How lengthy every of those duties is will range so much. Typically describing the duty to an agent and making a plan takes probably the most period of time, however in my expertise, the third step is commonly a time-consuming step, particularly for bigger implementations comparable to implementing a brand new characteristic or fixing a posh bug.
Now, to turn out to be extra environment friendly, what we need to do is to get rid of bottlenecks. The third step is a bottleneck we are able to simply reduce or get rid of. Steps one and two are very troublesome to be extra environment friendly at. It’s worthwhile to do that sequentially and spend time on them. Step 4 is certainly one thing you may make extra environment friendly, as I’ve described in different articles, comparable to my article on Making Claude Better at One-Shotting Implementations. I’ll thus give attention to step 3 on this article and how one can make it extra environment friendly.
The best way to run brokers in parallel
The straightforward answer to creating step 3 extra environment friendly is to run brokers in parallel. This sounds easy in principle, but it surely’s truly tougher to do successfully in apply. The explanation for that is that operating duties in parallel presents various challenges.
One problem it’s important to take care of is brokers writing code over one another. You don’t need brokers competing for enhancing a particular file or testing their very own implementation. So that is one downside it’s important to take care of. One other downside it’s important to take care of is the truth that it’s important to do context switching. So whenever you run duties in parallel, you naturally need to do context switching between the totally different duties. One facet of context switching is that it’s important to first describe one process and how one can resolve it, then describe one other process and how one can resolve it. One other facet of context switching is when the agent asks for suggestions, comparable to asking you questions on an implementation or asking you to check an implementation. So, minimizing context switching is one other facet that I’ll describe within the latter sections.
Working a number of brokers in the identical repository utilizing worktrees
The primary subject I’ll cowl is how one can run a number of brokers in the identical repository. The straightforward reply to how to do that is to make use of worktrees.
You may merely instruct your agent to make use of a worktree, which is a git characteristic that basically copies your total repository. So totally different brokers can work in utterly totally different repositories. After an agent finishes its implementation, it may be merged into the primary department of the primary repository, and on this approach, you’ll be able to have a number of brokers working in the identical repository with out conflicts.
Nevertheless, generally there could possibly be less complicated methods to attain the worktree setup. From my private expertise, I had points when utilizing Claude Code and asking it to take a look at a worktree. Claude would typically neglect to really take a look at a brand new worktree after which begin work in the primary repository. Then, when a number of brokers did the identical factor, I began having conflicts. Fortunately, Claude Code applied a –worktree command that you need to use when spinning up cloud code, comparable to beneath
claude --worktreeShould you use this command inside a GitHub repository, Claude will robotically take a look at to a brand new worktree given the present repository and can retailer all of the worktrees beneath a hidden Claude folder. That is extremely helpful for 2 important causes:
- You assure that you just’re all the time utilizing a worktree and that your brokers won’t intrude with one another.
- The worktrees are saved in a hidden Claude folder and never in your important folder. This considerably reduces the quantity of noise in your folder construction, making it simpler to navigate your folder construction manually if it’s essential to do this.
Should you’re utilizing one other coding agent, you’ll be able to work out how they’ve arrange worktrees, as worktrees have gotten an increasing number of of an anticipated characteristic for any coding agent program, comparable to an IDE like Cursor or a CLI like Claude Code.
Decrease context switching
Facet quantity two that it’s essential to take into consideration when operating coding brokers in parallel is that it’s essential to reduce context switching. I discussed earlier how this comes primarily in two features. One, when beginning after duties, and two, when interacting with the agent throughout the process, comparable to when the agent asks you questions, or it finishes its implementation and asks you to check it.
There’s no easy answer to simply get rid of contact switching, as parallel work inherently implies contact switching. Nevertheless, there are some things I consider to reduce my contact switching.
The very first thing I consider is to all the time end my present process earlier than shifting to a different process. This appears very apparent after I describe it; nonetheless, I need to present a particular instance. You will discover the instance described within the textual content beneath, however I’ll additionally present the picture to offer you a neater understanding of the idea proper beneath.
Let’s say you’ve gotten two duties, A and B, that it’s essential to do. Process A requires a five-minute interplay from the person, whereas process B requires a two-minute interplay by the person, and after that interplay, the agent will run for ten minutes to finish the duty. Additionally, you already began process A, and your psychological context is presently on process A.
On this state of affairs, it is rather tempting to change from process A to process B, do the two-minute setup on process B so the agent can run whilst you end process A. Nevertheless, I urge you to keep away from this as a lot as attainable and somewhat attempt to end the work you’re presently doing after which transfer on to the following process in line, regardless that this may not look like the optimum strategy at first look.
The explanation for that is which you can, after all, see what number of minutes it takes to finish every process. Process A takes 5 minutes, process B takes 2 minutes of interplay after which 10 minutes to run. Nevertheless, what you don’t see is the hidden price of context switching, and I’d argue that switching from process A to B after which switching again to process A will price you extra time than merely finishing process A after which finishing process B.
The second approach I reduce contact switching is to reduce distractions on my pc. There are quite a few methods to do that, however basically, I urge you to disable all notifications. I, for instance, disabled the Slack notifications on my pc and in addition the quantity showing on the Slack software that informs you in case you have a message ready for you. I discover this very distracting and removes my focus and context from the work that I’m presently doing.
Moreover, I attempt to hold one tab in my terminal setup per repository I’m engaged on. And if I’m engaged on a number of duties inside a single repository, I cut up the tab. I’ll focus on my terminal setup in additional element within the subsequent part.
Preserve an summary of your entire brokers
If you run a number of brokers, it’s extremely vital to maintain an summary of all of your brokers. Nevertheless, this doesn’t essentially come naturally, and it’s essential to make energetic selections to make it as simple as attainable to achieve an summary.
That is the place my terminal setup is available in clutch. When I’ve 5 to 10 brokers operating in parallel, it might probably simply be very complicated which agent is doing what and what it’s essential to work together with.
For my terminal setup, I exploit Warp. I believe Warp is an effective AI terminal that has good autocomplete and that I can simply work together with to make use of Claude Code successfully. I exploit one tab in Warp per repository I’m working in. This could possibly be a coding repository, for instance, containing the frontend or software, but it surely may be a gross sales repository the place I’ll do the entire gross sales work.
Then, if I run a number of brokers inside one repository, which I typically do, I merely cut up the tabs utilizing CMD + D on Mac. So I get a number of cut up subtabs for every of my important tabs. This helps me hold solely a single important tab for every repository, whereas additionally operating a number of brokers inside the repository.
Moreover, I wish to rename my tabs inside Warp to the repository identify in order that it’s simple to navigate to the right tab. And use CMD 1/2/3 and so forth to shortly navigate between the totally different tabs. And I’ve notifications from Warp at any time when an agent wants interplay.
I’d additionally like to notice that there are lots of different setups you need to use to your coding brokers. To begin with, you need to use a distinct terminal, which can doubtless work fairly nicely additionally, although my desire stays on heat. One other good various I do know of is to make use of the Claude software or an software known as Conductor, which makes it simple to offer an summary of your totally different repositories and coding brokers that you’re operating.
All in all, nonetheless, your setup comes right down to what your desire is. Should you discover a setup that works nicely for you, I extremely advocate utilizing that setup and easily sticking with it.
Conclusion
On this article, I’ve mentioned how one can effectively parallelize your programming duties. The primary purpose it is best to do that is to easily be extra environment friendly as an engineer. Should you run duties in parallel, you’ll be far more practical than an identical engineer who runs duties sequentially. Nevertheless, operating duties in parallel isn’t easy, and it’s essential to use particular methods to maintain an summary of all of your brokers, reduce context switching, and guarantee your brokers aren’t colliding with one another. I consider the idea I’ve described on this article is the way forward for work with AI. People will turn out to be orchestrators of AI brokers, and also you want to have the ability to effectively parallelize your work and fireplace off brokers to finish particular duties whilst you solely work together with the brokers on demand.
👉 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
👉 Discover me on socials:
💌 Substack
