(2025-03-22) Yegge Revenge Of The Junior Developer

Steve Yegge: Revenge of the junior developer. Brief note about the meaning of "vibe coding": In this post, I assume that vibe coding will grow up and people will use it for real engineering, with the "turn your brain off" version of it sticking around just for prototyping and fun projects.

For me, vibe coding just means letting the AI do the work. How closely you choose to pay attention to the AI's work depends solely on the problem at hand. For production, you pay attention; for prototypes, you chill. Either way, it’s vibe coding if you didn’t write it by hand.

Part 1: The six waves

Vibe coding is a whimsical name for chat-based coding, where you ask the LLM to write code, and then you feed it the results and ask it for more, in a continuous loop. It’s very different from traditional coding, or even coding with code completions.

Chat coding has been around a while in coding assistants, but without a rallying cry. It finally got one, when in early February the illustrious Dr. Andrej Karpathy, famed among other things for co-founding OpenAI, put a pretty name to the face of chat. He called it “vibe coding”

a strange, unprecedented, quantum-like triplet-state:

Vibe coding is still completely invisible to 80% of the industry outside Silicon Valley

Vibe coding is currently going batshit viral

Everyone’s still arguing about what “vibe coding” even means. But a ton of people, more every day, think it’s the future right now.

Chat coding in general is already utterly passé to an exponentially-faster growing group of developers who now wouldn’t walk across the street to piss on chat-based coding if it was on fire. They are still vibe coding and indeed getting better vibes than anyone. And yet they could not possibly care any less about your vibe coding

what you think of as vibe coding, and what I used to call CHOP – is indeed also still on the rise… for now. But agentic coding – the subject of this post – will soon rocket by chat coding as if it’s standing still.

Figure 1: Overlapping waves of AI coding modalities

The chart in Figure 1 depicts six overlapping waves of programming: traditional (2022), completions-based (2023), chat-based (2024), coding agents (2025 H1), agent clusters (2025 H2), and agent fleets (2026).

the figure depicts vibe coding as also increasing exponentially, but on a dotted line alongside the others, because as we’ll see in a bit, vibe coding is not a modality.

As a sneak preview for our discussion, "agent clusters" is the placeholder term I'm using for devs being able to run and fruitfully manage many coding agents in parallel

We all predicted the “wave four” coding agents were coming, and they arrived faster than most of us expected. And it's already possible to jump on wave five manually, albeit with effor

How do agents help with agents? Today, you have to notice when an agent worker is stuck, spinning, or done, and nudge it appropriately. Supervisor agents can and will start doing most of that for us very soon. The result: wave six.

Part 2: Where are you?

Code completions were very popular a year ago, a time that now feels like a distant prequel. But they are now the AI equivalent of “dead man walking.”

If you’re a bit more avant-garde, then you might think that chat-based programming is how things are going to unfold this year – meaning, in-IDE coding assistant chat interfaces like Copilot, Cursor, Sourcegraph, and Windsurf. If you’re in this group, then you're not doing too badly at all. Middle-of-the-pack, pat on the back.

But suddenly we have this latest wave, the new coding agents like Aider.chat and Claude Code – and soon, similar and prettier agents in all your favorite IDEs, wink nudge cough cough.

And the great thing is, with agents you are still vibe coding. That’s why it’s not a modality: You can vibe code with any non-manual AI modality: chat, agents, clusters.

Each successive modality wave, beginning with chat, is conservatively about 5x as productive as the previous wave.

Vibe coding remains a durable and lasting feature of that landscape – but not in the way most people think. Vibe coding simply means never writing code again.

Part 3: New camel owner’s manual

Let's examine how these new coding agents work, to help you see why this mere weeks-old development could quickly put your company in a real bind

Download and try out a coding agent, ideally one launched after March 1st 2025.

with agents, you don’t have to do all the ugly toil of bidirectional copy/paste and associated prompting, which is the slow human-y part. Instead, the agent takes over and handles that for you, only returning to chat with you when it finishes or gets stuck or you run out of cash.

they often get pretty darn far, entirely unassisted. They just grind away at their task until they get it right, throwing tokens at the problem to explore the space as needed

The only practical difference from chat, aside from cost, is that agents can perform much larger subtasks at a time, potentially encompassing many individual steps

you might tell a coding agent something like, "Here is JIRA ticket #; please go fix it." That is all you would need to say.

These new coding agents can solve huge issues, create even bigger messes, and generally behave like a supernaturally fast human developer who’s always flying a little blind and a little behind schedule.

Task graph decomposition, a skill we've all learned during the chat days of yore (December), is just as important today as you switch to vibe coding with agents. Even more so, because it’s so easy to overshoot and be over-ambitious with agents

Part 4: I was told there would be no math

This section is for CIOs and finance folks

how much opex budget did you put aside for developer LLM spend?

One company told me they were considering a generous budget of $25 per developer per day. That seems bold, like a lot of money. An almost reckless amount.
Well, it turns out they were on the right track. Coding agents are très cher, muy caro, we're talking very, very expensive. They burn lots of LLM tokens, to the tune of $10-$12/hour at current rates.

So it's going to be worth your while to budget more like $80-$100 of LLM spend per developer, per day

The upcoming wave, which I'm calling "agent clusters" – the chariot I hinted at in the last section – should make landfall by Q3.

Each of your devs will suddenly become like many devs

Agent clusters will have the side effect of finally moving software development into the cloud

Your dev desktop does not have enough power to run dozens of agents at once, let alone hundreds

Running N agents at once multiplies your devs' daily agent spend of $10/hr by N, not counting cloud costs, just token burn.

Part 5: The agent fleet is coming

Here's where it starts to get a little bit uncomfortable

The wave after clusters, or agent "fleets", for lack of a better word, will allow your developers to run 100+ agents at a time… with the help of yet more agents.

The new job of a software developer going forward will soon be managing dashboards of coding agents and their AI supervisors, as sketched in Figure 2: FY26 Org Chart

For you CIO-types, fleets will enable your developers to spend thousands of dollars a day. Even if inference costs plummet, the Jevons Paradox will result in higher usage offsetting those costs. If you don’t believe that, go ask to see your bug backlog; it’s basically infinite.

The scary part of this story is that if you can’t find or raise money, and you want to stay competitive, then you're going to have to make painful cuts in order to free up the opex budget. And if you run the numbers again, there's only one department where it makes sense to cut.

It turns out, it’s not all doom and gloom ahead. Far from it! There will be a bunch of jobs in the software industry. Just not the kind that involve writing code by hand like some sort of barbarian.

junior developers have actually been far more eager to adopt AI than senior devs. (contra (2024-12-10) Yegge The Death Of The Stubborn Developer)

They grab the O’Reilly AI Engineering book, which all devs need to know cover to cover now, and they treat it as job training

Whereas senior developers are, well… struggling, to put it gently.

They equated having to learn something new – and I mean really new, sort of like starting over

Not only are junior devs on average adopting AI faster, but junior devs are also – surprise! – cheaper. If companies are going to make cuts to pay for their devs to win with tokens, which devs do you think they’re gonna keep?

Agents are coming. Vast fleets of them. Not just coding agents. Agents are arising everywhere, across entire businesses and production tech processes. I talked to a big customer this morning who has already built dozens to hundreds of "AI task machines" – custom-built agents that perform specific parts of their giant workflows.

Switch to chat. Ditch completions.