A few days ago I saw a tweet from Brian Armstrong, CEO of Coinbase, announcing they were cutting 14% of their workforce. What caught everyone's attention wasn't the layoffs themselves but the rest of the text, where he explains how Coinbase is going to operate from now on under what he calls an AI-native model, with five layers maximum below the CEO, no pure managers, and small teams running fleets of AI agents. A few months ago I wrote on this blog about how agencies were going to split into three groups depending on how they responded to AI, and one of the hypotheses I laid out there was that the sector would shift toward fewer but better-paid jobs, with AI-augmented seniors getting most of the work and much less room for juniors. Brian's tweet is the public confirmation of that hypothesis, and that's why I wanted to share my take.
Cloudflare ran the same play forty-eight hours later
Firing people is one thing. Using a tweet with over 23 million views to publicly broadcast the operating manual you're going to apply from now on is something else entirely. What Brian was really doing was giving the rest of the sector permission to do the same, and it didn't take long for that to play out. Two days later, on May 7th, Cloudflare sent a letter to their employees announcing 1,100 layoffs, and the founders explicitly said it wasn't cost-cutting or performance management but a redefinition of how a world-class company operates in the age of agentic AI. The framing was practically identical to Coinbase's, and when two top-tier companies publish the same manifesto within forty-eight hours, the coincidence stops looking like a coincidence.
I work as a senior freelance Drupal developer, so this move doesn't hit me the same way it would hit someone inside Coinbase, though the second-order effects do reach me. And from where I'm standing, after a year and change wrestling with AI-driven workflows, I can say the technical part of the blueprint lines up with what I'm already seeing work.
How I work now
Day to day, I use a set of agents and automations I've built up through trial and error, and that lets me ship projects faster and with fewer silly mistakes along the way. Part of that setup is public on my repos, specifically a multi-agent system on top of DDEV built for working with Drupal, which I've been writing about here on the blog.
There's a clear cost to it, which is the token consumption, but in exchange I can iterate much faster and that more than makes up for it. What this approach doesn't save you is the work of reviewing what the AI is producing and deciding whether what it did is good or needs to be redirected. That review layer is where the whole thing gets decided, and it's exactly what Coinbase's blueprint takes for granted without explaining.
Judgment is what AI doesn't give you
Having judgment is what decides whether AI is genuinely helping you or just speeding up the rate at which you generate problems. AI writes code that often looks fine but is broken, looks broken but is actually fine, or just invents functions that don't exist. Catching that on the fly isn't something you learn from reading documentation or watching courses. You learn it by spending years on real projects, watching things break, and having someone with more experience next to you pointing out what you still can't see.
A junior doesn't have judgment yet because they don't have the background for it. That's not a criticism, it's just how the learning curve works. For simple tasks that isn't a problem because AI handles that kind of work perfectly well, which is exactly why it's replacing juniors on those tasks. The problem shows up on complex tasks, where you still need prior experience to know what architecture fits, what risks you're taking with each decision, and what to push back on when the agent doesn't deliver what you actually wanted.
Where are the seniors of 2036 going to come from?
If the model is that one generalist with judgment and a command of agentic AI does the work of a team, there's a question Brian's text doesn't answer: where are those generalists going to come from ten years from now?
The traditional career path worked through a silent deal. You started out doing the boring work nobody wanted to touch, you got paid less because you were learning from the seniors, and in exchange you absorbed how the bigger work got done. That deal is exactly what AI has broken, because if the agents do the boring work, the currency juniors used to pay for their training has lost its value, and if the agents do that work just as well or better, companies don't have a reason to hire them anymore. There's data from Stanford and the U.S. Bureau of Labor Statistics showing close to a 50% drop in junior hires between 2023 and 2025, and on LinkedIn the entry-level openings asking for three to five years of experience are already a meme. The generation entering the market right now is the first one that has to show up senior before it's allowed to be junior.
What a junior can do today
If companies aren't going to hire you to give you the experience you need, you have to go find that experience yourself. This isn't new, I've been recommending it on other posts since before the generative AI explosion: build side projects and ship them to production.
The typical portfolio with a static site and a contact form doesn't cut it. I'm talking about real projects, complex enough that they force you to learn new things while you build them. They can be yours, for small clients, or open source contributions to a module or a library you actually use. What matters is that they have some complexity, that they're live in production, and that you can show them off. On one hand you learn what nobody is going to teach you anymore, and on the other you've got something concrete in your portfolio that proves you can take a project from idea to deploy. And if you build them using AI agents, you're also gaining experience in exactly what the market is going to ask for, learning the limits, the traps, and your own best practices for avoiding problems in production.
How it looks from the freelance side
The quiet winner of this whole shift is the freelancer who works alone. Five years ago, getting a halfway serious project out the door needed a team with frontend, backend, design, copy, and someone coordinating it all. Today, with a decent agent setup, one person can take the whole thing on. The agent does the bulk of the work across each discipline and the person handles reviewing it with judgment so that what reaches the client matches the level they paid for. You don't need to be a deep expert in every area, you just need to know enough to judge what the AI hands back.
That fits with something else I've been seeing for months now. How I worked a year ago doesn't look like how I worked six months ago, and neither one looks much like how I work today. When this whole thing started, the models were assistants that completed your code while you were typing it, and today they're agents with quite a bit of autonomy that you hand a task to, let run, and review afterwards. Anyone who's hungry to learn and willing to put projects out there can catch up fast. The other way around, a senior with years of experience who refuses to use AI or to learn how to work with it better is going to find themselves out of the game sooner than they think.
There's a phrase that's been going around the sector for a while now and that strikes me as truer every day: we'll have work as long as clients don't know how to explain what they want. I walk into meetings where the client brings something like a letter to Santa and isn't able to see the implications of what they're asking for, or whether there's a simpler solution for what they actually need. My job there isn't to write code, it's to keep tugging until I understand what they're really asking for, translate that into a sensible set of requirements, and steer the project from idea to production making sure what gets delivered is maintainable and not overengineered. That role mixes translation, project management, and technical judgment, and it's what AI still can't do across a table from a client.
I don't fully know how all of this is going to land, but one thing seems pretty clear to me. The pure programmer who took orders and implemented them like a machine is the profile AI is replacing. The one who's pivoted into being a generalist, with the judgment to assess what AI puts out and the ability to actually understand the client, I don't see AI replacing that role anytime soon.