Embracing the parallel coding agent way of life

Embracing the parallel coding agent way of life

fifth October 2025

For some time now I’ve been listening to from engineers who run a number of coding brokers without delay—firing up a number of Claude Code or Codex CLI cases on the identical time, generally in the identical repo, generally towards a number of checkouts or git worktrees.

I used to be fairly skeptical about this at first. AI-generated code must be reviewed, which implies the pure bottleneck on all of that is how briskly I can evaluation the outcomes. It’s powerful maintaining with only a single LLM given how briskly they’ll churn issues out, the place’s the profit from working greater than one after the other if it simply leaves me additional behind?

Despite my misgivings, over the previous few weeks I’ve seen myself quietly beginning to embrace the parallel coding agent way of life.

I can solely deal with reviewing and touchdown one important change at a time, however I’m discovering an growing variety of duties that may nonetheless be fired off in parallel with out including an excessive amount of cognitive overhead to my major work.

Here are some patterns I’ve discovered for making use of parallel brokers successfully.

Research for proof of ideas

The first class of duties I’ve been making use of this sample to is analysis.

Research duties reply questions or present suggestions with out making modifications to a undertaking that you simply plan to maintain.

Lots of software program initiatives begin with a proof of idea. Can Yjs be used to implement a easy collaborative be aware writing instrument with a Python backend? The libraries exist, however do they work if you wire them collectively?

Today’s coding brokers can construct a proof of idea with new libraries and resolve these sorts of primary questions. Libraries too new to be within the coaching knowledge? Doesn’t matter: inform them to checkout the repos for these new dependencies and skim the code to determine how one can use them.

How does that work once more?

If you want a reminder about how a portion of your current system works, fashionable “reasoning” LLMs can present an in depth, actionable reply in only a minute or two.

It doesn’t matter how giant your codebase is: coding brokers are extraordinarily efficient with instruments like grep and might comply with codepaths by way of dozens of various recordsdata if they should.

Ask them to make notes on the place your signed cookies are set and skim, or how your software makes use of subprocesses and threads, or which features of your JSON API aren’t but coated by your documentation.

These LLM-generated explanations are value stashing away someplace, as a result of they’ll make glorious context to stick into additional prompts sooner or later.

Small upkeep duties

Now we’re transferring on to code edits that we intend to maintain, albeit with very low-stakes. It turns on the market are numerous issues that basically simply require a bit of bit of additional cognitive overhead which may be outsourced to a bot.

Warnings are an incredible instance. Is your check suite spitting out a warning that one thing you might be utilizing is deprecated? Chuck that at a bot—inform it to run the check suite and determine how one can repair the warning. No have to take a break from what you’re doing to resolve minor irritations like that.

There is a particular knack to recognizing alternatives like this. As all the time, one of the best ways to develop that intuition is to strive issues—any small upkeep activity is one thing that’s value making an attempt with a coding agent. You can study from each their successes and their failures.

Carefully specified and directed precise work

Reviewing code that lands in your desk out of nowhere is a lot of labor. First you must derive the objectives of the brand new implementation: what’s it making an attempt to realize? Is this one thing the undertaking wants? Is the strategy taken the most effective for this present undertaking, given different future deliberate adjustments? Lots of large questions earlier than you possibly can even begin digging into the small print of the code.

Code that began from your individual specification is quite a bit much less effort to evaluation. If you already determined what to unravel, picked the strategy and labored out an in depth specification for the work itself, confirming it was constructed to your wants can take quite a bit much less time.

I described my extra authoritarian strategy to prompting fashions for code again in March. If I inform them precisely how one can construct one thing the work wanted to evaluation the ensuing adjustments is an entire lot much less taxing.

How I’m utilizing these instruments at the moment

My each day drivers are at the moment Claude Code (on Sonnet 4.5), Codex CLI (on GPT-5-Codex), and Codex Cloud (for asynchronous duties, steadily launched from my telephone.)

I’m additionally dabbling with GitHub Copilot Coding Agent (the agent baked into the GitHub.com internet interface in varied locations) and Google Jules, Google’s currently-free different to Codex Cloud.

I’m nonetheless settling into patterns that work for me. I think about I’ll be iterating on my processes for a very long time to come back, particularly because the panorama of coding brokers continues to evolve.

I steadily have a number of terminal home windows open working totally different coding brokers in numerous directories. These are at the moment a mix of Claude Code and Codex CLI, working in YOLO mode (no approvals) for duties the place I’m assured malicious directions can’t sneak into the context.

(I would like to begin habitually working my native brokers in Docker containers to additional restrict the blast radius if one thing goes incorrect.)

I haven’t adopted git worktrees but: if I wish to run two brokers in isolation towards the identical repo I do a recent checkout, usually into /tmp.

For riskier duties I’m at the moment utilizing asynchronous coding brokers—normally Codex Cloud—so if something goes incorrect the worst that may occur is my supply code getting leaked (since I enable it to have community entry whereas working). Most of what I work on is open supply anyway in order that’s not a giant concern for me.

I often use GitHub Codespaces to run VS Code’s agent mode, which is surprisingly efficient and runs straight in my browser. This is especially nice for workshops and demos since it really works for anybody with GitHub account, no additional API key crucial.

Please share your patterns that work

This class of coding agent software program remains to be actually new, and the fashions have solely actually obtained adequate to drive them successfully up to now few months—Claude 4 and GPT-5 specifically.

I plan to jot down extra as I determine the methods of utilizing them which can be simplest. I encourage different practitioners to do the identical!