Contrary to what your feed wants you to believe, the biggest productivity unlock this year isn’t parallel agents. It’s the ubiquitous dev environment. No affiliations.
The 1000x engineer
It’s almost half-time 2026, and the only thing holding your portfolio back from hitting a life-altering exponential is a single question, nagging at the back of every stock market analyst’s balding head: is it a bubble, or are there actual productivity gains?
Of course there are. If anyone would know, it would be us: the expendable coders. And not because of the never-ending tech layoffs (by now it should be clear to everyone that there are entirely different motives for that). No, the reason is because every day we see our computers do things for us that just a few years ago would have seemed like magic. They write 100% of our code. They test it. They review it. Of course, there are productivity gains.
How much?! the analyst yells.
My goodman. Developers couldn’t even estimate effort before AI. What makes you think we can now?
The gains are real. Period. And yet, there’s something strange about the current narrative. Everywhere I look, I see the same scene; a powerful programmer, hunched in front of a wall of monitors. In some sorcerous trance, he expertly wields his keyboard, orchestrating dozens of agents, each with their own tmux session, git worktree and branch. When he calls it a day, he leaves the agents running overnight. All through the night they grind, producing perfectly secure, efficient, bug-free, scalable, testable, readable code. Just in time for his morning coffee, they deliver dozens of perfectly scoped little PRs ready for his well-rested eyes to review.
Is he real? Did the renowned 10x engineer just become 100x? 1000x? A mill…?
Narrative matters. Especially at a time like this. At the dawn of the fourth industrial revolution. On the verge of the biggest technological shift ever. At the last gasp of the current western societal contract. It can be the difference between a supercycle and a 1929-scale crash. Damnit, this is your portfolio we’re talking about!
To the balding analysts, I don’t blame you. You have every right to call out hyperbolic claims like these. Because that’s not how software engineering works. Not in production. Definitely not at scale. Let’s back up.
AI is eating Agile
The pre-AI development cycle used to look something like this:
You plan, you build, you test, you review, and then you ship. Simple. Then Cursor came along, and with it the first somewhat helpful coding agent. Though neither the models or the harness were good enough to fully delegate a whole cycle step to, they relieved us of a significant share of our coding and testing. It was the good old copilot era, back when we could still laugh at the notion that these silly LLMs could ever compete with us.
Then no one was laughing anymore. Claude Code, opus-4.5, skills and Ralph appeared, and in the span of a few months us humans went from writing ~50% of our code to 0. Finally, we had a harness and a model that were actually good enough to fully delegate the build and test steps to (and, partly, planning and reviewing). The idea behind @GeoffreyHuntley’s Ralph framework was as simple as it was beautiful: wrap the coding agent loop in another loop. Suddenly, we no longer gave our agents tasks. We gave them sprints:
And now, in the spring of 2026, the 1000x engineer is trying to tell us that the models and harnesses have become so good that we no longer need to run our agents one task at a time. If you still do: NGMI. If you tried to go parallel, but couldn’t make it work without losing your mind: skill issue.
What they're actually saying is this:
The emperor’s new clothes
Here’s the thing: though AI will likely change everything about how we work, there’s still no replacement for the Agile dev cycle. Not if you believe human judgement still has a role to play in product development. Not unless you believe the planning and review steps can be automated.
When it comes to planning, most good engineers spend more time planning than ever. Why? Because the more time you spend planning, the less time you have to spend reviewing. Because the agents will actually understand what you want and not feel the need to improvise (@mattpocockuk's grill-me skills have done wonders to prove this point). With thorough planning, you’ll know what to look for in the review. If the PR is good, nothing will feel foreign to you.
When it comes to reviewing, it’s the other way around. We’d all rather spend as little time as possible doing it. It’s the worst part of software engineering. Always has been, always will be. So believe me when I say: there are few SWE domains with bigger automation incentives. And yet, here we are. After all this time, there’s still no decent solution. We’ve all seen the review bots. The skill packages and frameworks. The Ralph-on-steroids. At best, they’re nitpicking on things I actually agree with but that makes literally no difference to anyone but the most anal red-tapist. At worst, they give me walls of text without substance. And to this day, I’ve yet to see a review bot tell me a file shouldn’t exist. Or even better; that the PR shouldn’t.
So, where does that leave the 1000x engineer? Well, let’s be conservative. Let’s say he’s running 20 agents in parallel at all times. What does it mean for planning and reviewing?
Man plans, God laughs
Anyone who’s tried the grill-me skills knows that any decent model-harness combo will take you to hell and back. It’s literally torture. But unfortunately, a necessary torture to smash the human fallibility out of our silly, naive plans. A single of these interrogation sessions can easily run for 30 minutes. Ok, so if the 1000x engineer has 20 agents running in parallel, it could only mean one of two things. Either he’s automated the planning part, which means he’s letting his agents decide for him on the myriad unknowns and edge cases buried in his one-shot prompt. Or, he’s suffering through 20 of these torture sessions all at once. If it’s the latter, bless his sweet little heart.
Regarding the review step, assuming he actually cares about the code being deployed to his users, it means he’s got 20 PRs to review at any given time. And after every merge, he’ll need to rebase every single PR with a merge conflict. Congrats, you just DDoS'd your own context window.
There’s no way around it; parallel agents don’t scale. It’s a fugazi. A whazy. A woozie. Fairy dust. For those of us who’ve actually tried to make use of them in enterprise-grade systems, they’re hopelessly constrained by our own primitive, sequential wetware. Ergo, the 1000x engineer is either cheating, or he’s not sleeping. Neither will result in good software.
Hidden in plain sight
Acceptance is the first step to recovery. Only when you stop looking at the trees can you see the forest, and so on. Only when I realized parallel agents were a lie could I see the real productivity unlock of the past few months. The bottleneck was never in my methods. It was me. If I wanna up my productivity, the solution isn’t more agents. It’s more of me.
Think about it. It’s 2026. We all agree coding (as in the actual typing of the syntax) has been solved. @bcherny was right. So why are we pretending like it hasn’t? Why are we still chaining ourselves to our pre-AI desktop workstations for 10 hours a day? Clocking in on the factory floor like it’s 1850, as if machines hadn’t just automated our jobs.
The difference between now and then is that we weren’t replaced. We were promoted. We’re no longer workers on the assembly line; we’re managers. As such, it’s no longer our job to make the thing. It’s to facilitate those who do. To plan, provide the means of production, review, and—most important of all—support.
We’re no longer workers on the assembly line; we’re managers.
The beautiful thing is that we can do most of this from anywhere. For most software engineering tasks today, we don’t actually need a desk, keyboard, mouse, or a wall of monitors. We just need to be available for our agents.
I give you: the ubiquitous dev environment. Or rather, @levelsio, along with a bunch of other trailblazers, give it to us:
Heard of Claude Code Remote Control? the 1000x engineer blurts, frothing at the mouth. Codex remote connections?
Easy. No matter how much Anthropic and OpenAI would like us all to use more of their lock-in features, they’re both missing the point. What I’m talking about is your actual dev environment. Your machine. Your shell. Your editor. Cloud-hosted (or on-prem). Isolated. Vendor-agnostic. Secure. Seamlessly accessible from any device of your choosing. Your full-featured workstation, in the palm of your hand.
Math not mathing? Read this:
https://x.com/mag_sven/status/2060077593617899987?s=20
Even so, just because a lot of the work is mobile, it doesn’t mean we’re ready to go 100% mobile. Due to the planning and review bottlenecks, the harness still needs a human-in-the-loop. That means we need to understand the thing: to hold enough of the system in our heads to know when a diff should be rejected. Which means we need to be able to test it, run it, and deploy it. Which means the CC and Codex remote features aren’t enough. We still need the CLI agents. We still need our own dev env.
After trying @levelsio’s setup for a few weeks, I can honestly say it’s one of the most profound changes to my workflow ever (close second to Claude Code). The pure bliss that is shutting your laptop and picking up right where you and your CLI agent left off on your phone is something that has to be experienced. Enable voice, and you’re officially a time-traveler.
It’s also made me realize just how much time I spend every day in an idle state. Scrolling in bed. Riding the bus/subway. Waiting for people. Scrolling on the toilet. Walking the dog. Resting between sets at the gym. Scrolling on the couch. By freeing myself from the old-world workstation, I can actually work from anywhere. Of course, some of this time is best used as a break from work (this isn’t hustle porn, I promise). But, surprisingly, much less than I thought. Turns out, most of these activities are actually parallelizable.
What’s more, the ubiquitous dev env has made me fall back in love with tinkering. With the rapid progress in our field, it’s easy to get the sense that those days are over. In many ways, they are. But at least this project, I tell myself, is worth my time. Running a factory is essentially a communication problem, so surely perfecting my personal command center must be key? Every single day, my agents and I run into new problems. For every problem, we learn something new. And for every solution, a door opens to a new problem. Just like it always was. And that's the strange part: I'm running the most advanced AI workflow yet, and somehow I feel closer than ever to the programming I once knew.
Narrative matters. By not getting caught in the parallel agents hype trap, serious software engineers can maintain both their sanity and pre-AI levels of product quality. And, ironically, by adopting a ubiquitous dev environment, they can get way more done. Accept the fact that you’re the bottleneck, and structure your workflows accordingly. Remember, your job is no longer to make the thing. It’s to support those who do.
Parallelize yourself, not your agents.
Top comments (0)