The brute squad

Welcome back! Come one, come all, friends, foes, fart connoisseurs, all are welcome here at Camel Central. It has been an action-packed three months since Revenge of the Junior Developer (RotJD), which is essential reading for this post, so shoo, off you go. You might also want to watch The Princess Bride, up to you. As you wish!

What has changed since March? Much and little, more or less. For starters, models got better. Claude 3.7, every programmer's favorite, is now nearly four months old, one foot already in the grave, sounding like a talking parrot compared to the newer models. Recent releases from Anthropic, OpenAI and Google are chewing through problems that were unsolvable by their predecessors. At least on my box. We're still on those curves, right on schedule.

We've also got a crop of exciting autonomous-agent tools, the so-called camels that are replacing all that walking you're doing with Cursor right now. For instance, in response to the game-changing Claude Code, OpenAI launched their own coding agent, which they named Codex, in a clever nod to how nobody can understand any of their model names or when they came out or how they relate to each other or which one is better for any given task. Codex is ideal for anyone who understands that stuff, e.g. presumably certain OpenAI employees, possibly some of the models themselves.

Rounding out the trio, Sourcegraph launched Amp, which is a team-oriented enterprise coding agent, with leaderboards, public workstreams, and friendly competition. People are finding Amp refreshing and our users are effusive about it; even Swyx gave it a nod recently (yo Swyx!)

So now you have choices. Claude Code, Codex, and Amp: three serious commercial autonomous coding agent contenders. And there are also some great OSS agents improving fast, like Cline and RooCode, for which I have high hopes.

What else is new since March? I am relieved to report that I have at long last finished co-authoring a book on vibe coding, along with my buddy, famed writer/researcher Gene Kim. It was a months-long slog for us to make it an hours-long read. It's our instruction manual for vibe coding, both with chat and with autonomous agents. Vibe coding professionally is a complex craft, and a topic for another day. But I'll drop a link at the end, and I'll be posting more about it soon.

One other small bit of news, prompted in part by RotJD, is that in April I got to meet Dario Amodei, CEO of Anthropic, for a private half hour 1:1 chat at their HQ. Unfortunately on the way over, my brain received a remote firmware downgrade, leaving me with the intelligence and personality of a sea cucumber for approximately thirty-five minutes. But it was a delightful discussion, uh, for me, and I will share some of the highlights toward the end of the post.

And with that, on with the show! Our production today features more of everything you loved from RotJD. In this post you'll be treated to crazy news and crazier stories that touch on three themes: 1) Dev work is changing, 2) Knowledge work is changing, and 3) Society is changing. We'll have more predictions, more facts, and more flatulence; I leave it to you to decide what's what.

If you don't like long form reading, then, TL;DR: Either learn to use coding agents, or start learning a trade. I hear electricians drive the nicest cars on the construction sites. And plumbers are the new millionaire class. You've got options.

I prefer long form stories though, don't you? You laugh, you cry, you remember it forever. If you're down, let's go!

Part 1: The death of the IDE

Today's first exciting topic: all your favorite software development tools are getting killed off faster than a new Netflix series. What, you say? Who shot who in the what now?

Well, not exactly killed off per se. Dying off, more like. They're being replaced by agentic coding: a brand new way to develop software since February, exponentially gaining popularity all around you, though you may be blissfully oblivious. Agentic coding is short for "vibe coding with autonomous coding agents." The typical agentic coding workflow doesn't involve using traditional IDEs like, say, Cursor, or Windsurf. In the limit, you don't use them at all.

Instead, agentic coding is console based, 1970s-style, in a text terminal window with little to no UI. The developer workflow is simple: You ask the AI for code changes, then you review the diffs and test the output, in a loop, lather, rinse, repeat. It's a bit like you're directing a movie, but micromanaging the crew's work more obsessively than Stanley Kubrick. It can be both as fun and as stressful as directing a movie, but once you start, you never go back to the old way.

A buddy at Anthropic told me that engineers there rarely use their IDEs anymore. To agentic vibe coders, opening an IDE is like opening the Chrome developer tools, or maybe popping the hood on your Uber: an undesirable scenario, a move of last resort.

I mean, it's just one data point. One company-sized anecdote. I wouldn't be bothering you with such implausible-sounding accounts, except that it's happening to me too.

To wit: I got a new work computer a few weeks back. During setup, as is my sacred duty, I installed IntelliJ and VS Code, both IDEs that I love, and readied them for development with my keys, personal configuration, etc.

And then I promptly never opened them again.

Oh, I'm coding. I've been coding on the new macbook for weeks, having eye-twitching levels of fun. I'm producing more code per unit time than ever, and it's good code, too. But there is nary an IDE in sight—not even Emacs, except of course when I use it for its non-coding capabilities such as Ad-Hoc Dynamic Persistence Engine, Replacement Operating System, Friend, etc.

But I don't seem to be using Emacs to look at any code, which is out of character for me. Like, mustache-twirly out of character. What could be going on?

None of this was intentional. It's not as if I swore off IDEs. It's just that when I think about opening up an IDE and coding the traditional way, I would rather have to wash my car with my own saliva.

You will meet plenty of programmers who want to keep coding the old way. They'll insist that you'll never get your car truly spotless unless you spit-shine the whole thing like grandpaw did, and they'll lick everything clean from their windshield to their rear axle differential to drive their point home.

They are being ridiculous. Anyone can see that if you were to pit a traditional spit-washer in a heated battle of Car Cleaning against someone armed with, say, a car wash, then after a few hours there would be two clean cars, a dehydrated dude covered in grime who can no longer move his tongue, and you, who watched it all from the sideline with a Slurpee.

Whatever the car-lickers might think, there is no "competing" with coding agents, not even close. It's a brave new world, and I never think about going back. To IDEs, I mean. Remember we were talking about them? Each coding agent is a car wash. And I believe my job is to clean cars. Not lick them.

In the new world of software development, my job is to produce code, not write it. Let's leave the IDEs to, and for, the AIs.

Part 2: Vibe coding on the edge

Coding the New Way is, alas, unintuitive and trap-prone. Learning to use agents safely is a ______ process. Feel free to madlib adjectives there, I went with the f-word myself. But I'm getting better as the months go by, up to juggling four or five agents with their own workstreams at once, which feels like a sweet spot with today's tech.

I've recently embarked on a quest to be able to run all my agents while I play video games. And not just so I can get fragged by two kinds of teammates at once. No. I'm doing it as a distraction. Vibe coding with multiple agents can be intense. The last thing I need is to run even more of them.

Other activities, I can put down and walk away from. My agents, though, are a team of programming brutes standing around waiting for stuff to do. This makes them hard to ignore, since I have a lot of work. I feel like when I'm not using them, I'm wasting time and potential, because in theory they could be working around the clock. And by "in theory," I mean "if I somehow pulled off the biggest bank heist in modern history to pay for the tokens."

Somehow my agents have turned into my babies, programming-brute baby birds in a nest of console windows, squawking at me, mouths agape. Coding agents are nothing if not hungry. That's one of the reasons I tend to use Amp; I feel it squawks a bit less. Whenever a bird gets hungry, we sketch out new work, and then I send it on its birdy way, out into the wide world of my unprotected hard disk, network, and bank account.

I characterized coding agents as camels in RotJD, and that metaphor still has legs. Four of them. But I find it difficult, as a writer, to convey the siren-song allure of coding agents when I'm describing them as a bunch of camels gurgling at me. So they're… baby camels? Regardless, they're like toddlers, but with proper supervision and food preparation, they're absolute brutes at writing code.

I can't bring myself to leave when my brutes are hungry. I've tried, it's a no-go. I have to run a practiced escape plan every night to get my computer closed by 2am. First I get them all spinning at once. Then I leap up, run out of the room, slam the door, jam my fingers in my ears, and sprint away shouting lalalala. It is perhaps not the most refined of plans, and doesn't have a great success rate. My babies need me.

But consider! My hypothesis – you can now see the genius of my new project – is that I could instead be playing a video game around 2am, get killed and rage quit, and stomp out of the room before I remember my hungry camel bird whatevers. Once I'm at a safe distance, I could get some sleep, use the restroom, shower, all the stuff I should have done twenty hours before. Easy peasy.

That, friendos, is the story of why I want to play video games while I run four coding agents.

This turned out to be the biggest surprise of the new world: agentic coding is addictive. You will hear it more and more often, because it bewitches people once they've got the hang of it. Agentic coding is like a slot machine, where each of your requests is a pull of the lever with potentially infinite upside or downside. On any given query, you don't know if it's going to one-shot everything you wished for, or delete your repo and send weenie pics to your grandma.

Every time something good happens, which is often, you get rewarded with dopamine. And when something bad happens, also often, you get adrenaline. The intermittent reinforcement of those dopamine and adrenaline hits creates the core addictive pull. It can become near-impossible to tear yourself away. We had to drag several vibe coders off stage at a conference I was at recently. As we escorted them away from the podium, they would still be wailing, "It'll work on the next try!"

So be careful. It's potent stuff. If you do attempt running six agents in separate workstreams, bring potable water and a couple of empty jugs.

You've now had a peek into my weird new world of not thinking about IDEs anymore. AIs themselves are just beginning to take over the use of IDEs. It's unstoppable, because vibe coding is a lower energy state for producing software, so everyone who can will eventually migrate there. People are already figuring out how, and the supporting tooling and model improvements will continue to accelerate.

So there you have it. The agentic vibe coders are coming. If you spot someone with a bunch of colorful terminals open, no IDE anywhere to be seen, waterfalls of code scrolling past, arguing with the computer like a stereotypical Hollywood computer hacker – then you are watching someone who is already living in the future. And they really need to go to the bathroom. Do them a favor and slam their laptop shut. They will thank you for it, in their way.

Part 3: Rise of the vibe coder

I wouldn't be regaling you with any of this if we didn't see it playing out in other companies too. "We" meaning both Sourcegraph, and our customers, and Gene Kim, and his fellow researchers. We all hear the same story unfold. We're seeing it happen to ourselves, too.

It's not just software engineers that are picking up vibe coding with autonomous agents. I also have stories I could tell you from Product, Sales, Marketing, Finance, Customer Success, UX and more. We've seen vibe coding cropping up in all departments. I had a whole Part 4 written about it, Rise of the 2-Pizza Team, that I had to cut for space. Next time!

For our purposes, the news here is that vibe coding with agents is beginning to have non-ignorable consequences that are getting the attention of HR, Legal, Finance, and executive teams.

One such consequence is that these new programmers are badass. They're sending PRs at a rate 10x higher than their peers. And while we're hearing that more AI-submitted PRs are rejected in review, the number that make it through dwarf the work of the programmers who still use IDEs, even Cursor and Windsurf.

The agent/IDE disparity in productivity is becoming so pronounced that Performance Review time is looking to be a bloodbath. Some Eng leadership teams are beginning to have discussions with their Legal and HR teams about options. How do you meaningfully calibrate the employee who operates a car wash against the ones on Team Drool?

Some of these companies already want all their engineers vibe coding with agents, even though today the "car wash" is just a bunch of camels licking the vehicle. Doesn't matter, still cleans faster. Avant-garde companies wish their engineers would adopt it. But they're getting severe lickback from devs who are spittin' mad, hawking stories about how the company should just spit it out and admit their AI push is hard to swallow.

Companies are responding to this pushback with all the empathy of Mr. Burns mashing the Release the Hounds button. I'm sorry, just a journalist here, reporting what I see on the ground. Some companies, big ones, are telling us that they are beginning to consider "displacing" people who cannot or will not adopt agentic coding. In at least two cases, I've heard the company is thinking of firing up to 60% of their engineers, simply because they wouldn't switch. I warned of this in RotJD, and it's already begun.

Vibe coding makes engineers badass. But this in turn makes the non-vibe coders look bad. Terrible, even. That kind of situation at a company is unstable and never lasts for long.

How do you know if you're doing AI right at your company? We've noticed that the companies that are winning with AI – the ones happy with their progress – tend to be the ones that encourage token burn. Token spend per developer per unit time is the new health metric that best represents how well your company is doing with AI: an idea proposed by Dr. Matt Beane and playing out in the field as we speak. I see companies saying, "If our devs are spending $100-$300 a day, that's much less than paying for another human engineer. So if AI makes our devs twice as productive, or in some cases only 50% more, we're winning."

There's an implicit assumption here that tokens == productivity. We don't have any data to back this up yet, so if you're about to have a seizure from wanting to argue with me, this would be a good spot to pounce and call bullshit on every word of this post.

The token spend has another key benefit. Early enterprise vibe coding adopters are finding that when you remove coding as the bottleneck, the bottlenecks move elsewhere in the system, much like what happens in manufacturing. Previously undetected issues with your systems and business processes will emerge as urgent first-class blockers. If you learn about them early, you can fix them early. Sadly, going into detail would be another full post.

But if you're not sure what to do about your rising token spend, "Burn Baby Burn" is not a bad motto right now. Just don't burn the house down.

So! We've talked a little about how both coding and broader knowledge work are moving to vibe coding. Let's broaden our perspective by looking at some different visions of how all society might change. We'll see what they have in common, and talk about what to pack in your go-bag.

Part 4: The endgame (2026)

As I mentioned in the intro, I had the privilege of meeting Dario Amodei, co-founder and CEO of Anthropic, at their corp HQ in SF. It was unexpected and hastily arranged, but I jumped at the chance. How often do you get invited to a private 30-minute audience with a CEO who might one day be one of the most powerful people in the world? It felt like a once-in-a-lifetime opportunity, and in hindsight, I made it one.

You see there was a minor mishap. As far as I can recollect, I was on my way up to meet Dario, right outside his beautiful office, when a hole opened up in the ozone layer and I was bombarded by Dipshit Rays from Outer Space, which caused me to speak nothing but Dipshit until just after I left the building. You know those documentaries where surgeons poke around someone's brain with little cattle prods, and the person smells bacon, or sees a rabbit, or only speaks fluent French even though they never learned it? Well somewhere along the line, I learned to speak advanced Dipshit.

Anyway, it's been several weeks, and lacking other compelling hypotheses, I've settled on Dipshit Rays. I'm putting the matter aside lest I cut myself on Occam's Razor. Let's take a look at what I learned.

Back in October, Dario wrote what I think will become a historic essay, Machines of Loving Grace, in which he reasons through what may happen to society when we are no longer the smartest ones on the planet, which happens in about 2 years. It is a glass-half-full vision, in my opinion, both uplifting and thought-provoking. Our conversation ran along those lines.

I enjoyed our meeting immensely, despite having lost all control of my mouth, which frankly should not have been invited along. Dario was gracious, warm, animated, funny, personable, well-spoken, and good enough to share with me some of his less-varnished, more exploratory ideas around the impending future.

To me, the picture he painted seemed to lean a bit glass-half-emptier than what he shares with journalists and John Q. If I were to paraphrase, AI will change knowledge work and all jobs that use computers, raising up people who are good at working with AI, and displacing those who are not. That displacement – such a large-scale reshuffling of knowledge workers – could result in large-scale unrest. Dario didn't claim either vision to be more accurate; nobody can know.

One overarching impression pulsed in my crustacean brain throughout our chat: Dario cares deeply about what is going to happen to us. He is one of the best-informed people in the world on this topic, and sees things changing for everyone soon. As to how, Dario exudes equal amounts of hope and concern, and I think that balance would be good for all of us.

Dario's private glass-half-empty thoughts were comparatively mild, anticipating some socioeconomic impact, social unrest, and maybe some players or regions gaining insurmountable leads. For a glass-destroyed-by-robots vision of the future, AI 2027 offers a grim choose-your-own-adventure in which all roads end in about two years with AIs taking over the world. It's a good story. I read it as I was kicking myself on the flight home. In the path I picked, I was living in the forest and got mopped up by drones, which at the time didn't sound so bad. But it failed to lighten my mood, because it was unrealistic.

There are many problems with AI 2027, but most notably its over-optimistic schedule. It may well happen someday. But nation-state actors are slow to act even when they're trying to be fast. And supply chains, already stressed to their limits, will become hard bottlenecks. Even superhuman intelligence will run aground on bureaucracy, energy limits, and other real-world obstacles.

No matter whose vision you subscribe to, something big will happen with AI in the next two years. Compute power for AI is doubling around every 100 days, which makes it likely that we hit AGI by 2028 if not sooner. But I take an optimistic view: Our problems will always be harder than anyone can fully understand, even superhuman intelligences. It will be nice to have them around to help us solve those giant problems.

My biggest takeaway from my meeting with Dario is that he anticipates a big mess next year. He casually refers to 2026 as "The Endgame" without a hint of drama, as if we were talking about a movie we're watching. He characterizes society as the classic immovable object, and tech as the unstoppable force, except this time it's not a thought exercise. Tech is about to push society harder than society wants to be pushed – harder than society is willing to be pushed. He smacks his fist into his palm very convincingly when he says "willing".

I left with the impression Dario feels that the collision of tech and society, despite being mostly helpful and awesome, will cause some shockwaves. Toward the end he even pondered whether the rest of the world will be able to keep up with the Bay Area during the dramatic expansion and adoption of AI, since outside the core tech bubble, people seem to be slower to embrace it.

Whatever happens, most of it will be really, really cool.

Part 5: Why you should be excited

I promised I wouldn't talk much about our book on Vibe Coding. I will provide a link at the end of this post. It won't be out until September, anyway. The book, not the end of the post. Startin' to feel that way though, isn't it? I agree. Let's wrap it up.

Programming is changing. Knowledge work is changing. These are exciting times. Those who embrace the change will thrive. They are already thriving.

Traditional programming is Vizzini. Clever, but weak. With Cursor/Windsurf you get upgraded to Inigo Montoya. Great fighter, fast, lots of fun. But working with a coding agent, you become the giant Fezzik—except also as fast as Inigo and as smart as Vizzini. A true coding brute. Even so, despite Miracle Max's joke, you aren't a squad yet. Even Fezzik said he's just on it.

But if you're running multiple coding agents at once, four or five Fezziks working for you? Then you ARE the Brute Squad.

Agentic coding lets you be faster, more ambitious, and even have more fun. But it's not easy! Your car wash can easily switch into sandpaper and tiger tongues. Agentic vibe coding is powerful, but requires new kinds of supervision. That's why you need to start learning it now. It takes time, and practice, to feed those birdies and keep them from chewing through the wiring.

Is it worth it? Hell yes. All I can tell you is that I was not expecting any of this. All my initial biases and preconceptions were wrong. I scoffed when I first heard about terminal-based coding back in January. It sounded ridiculous. But I am having more fun with agentic coding than I had with any other kind, at any time in my career. This is better than the absolute best times I had building my game, launching cool stuff at big companies, and making open-source contributions like js2-mode. Those projects were all so fun they would give me Writer's High. You know the feeling.

Agentic coding is every bit as fun as they were, but concentrated and accelerated, with a pretty big awe/wow factor thrown in. It's like sticking your head out the car window at full speed. But you're also driving, so be careful! After you get the hang of it, it becomes more like driving a convertible – exhilarating, but more refined. And if you're like me, and disable all permission checks on your agents, it's like driving your convertible with no seatbelt, wearing a kilt, only using the parking brake, and flooring it everywhere. It's a bit of an acquired taste. Never goin' back though.

If you're still skeptical, go watch the underrated movie Pleasantville (1998), which starts in monochrome and gradually people turn color. Solid movie, and a metaphor for many things, but none more appropriate than how coding is changing all around you. You are clinging to a vanishing past. One you won't even miss. Why wait?

Switch to a coding agent! It's time. Claude Code and Codex are both great, and if you use either one, then power to you. Same if you're using Cline/Roo. You're a badass, you're on the right track, you're doing the right things, the future is bright for you. Be the Brute Squad!

Or you could just use Amp. I can use any agents I like, and I wind up using Sourcegraph Amp for most of my work. Amp is like the others but with a couple edges. One is that it annoys me less than the other two. Even a 2% difference in Annoyance Factor can gradually push you towards a lower-friction solution, and I find Amp to be a more reliable workhorse. I pull the other agents out only when Amp gets stuck. (They can all get stuck; it's good to have more than one available. All this and more is in our book!)

Amp is also more fun. It takes a different design approach, being intentionally team-centric. Amp gamifies your agentic development by making it public, with leaderboards and friendly competition, as well as liberal thread sharing. It all manages to be low-pressure and engaging. Who would have guessed that coding was such a community activity? But here we are.

Professional vibe coding is growing complex. Gene Kim and I each made many attempts to share what we have learned. Gene and I have been vibe coding since before Dr. Karpathy gave it a name, and we've learned individually, together, and from (and with) others. But every tutorial turned into a 3-hour video, a 30-page blog post, or a 3000-word tweet. It is just too much information to get across and organize in a social media post.

So we decided we had to write a book. Ét voila, five months and 380 pages later: Vibe Coding. It was a lot of work to make it a fast read, but I think we managed: you can get through it in a few hours. I've reread the book a couple times, always dreading it, and it always turns out pretty darn good. I think you'll learn a thing or two from it. I always do.

Thanks for reading to the end, it's been fun. Actually it's been terrible, months in the making, I rewrote this a dozen times. But my thanks goes out to all the reviewers who critiqued today's post, especially Dominic Cooney and Dr. Matt Beane.

Dominic is calling it the Second Renaissance, and I think he's right. You could pick any number of examples, but look at photos, which have gone from infinite cost circa 1875 to infinite wealth today; photos are free and everyone takes them all day. Now those photos are becoming like the talking pictures at Hogwarts. Everyone will create bespoke software using AI. It will change everything forever.

And engineers, being trained, will always be the best at it. Building enterprise software will always be monumentally difficult, so engineers and AIs will team up to build it together.

Hop on the bandwagon, folks! It's just taking off. You haven't missed out on anything but bugs so far. The coding agents and best practices are all getting smoother. You can learn them, and soon you will love them.

See you next time! I am off to have a stern talk with my mouth, and tell it to shut the hell up.