The staggering economics of agentic coders
Who's going to pay a human dev $150k per year when the equivalent AI dev is just $62.50?
I’ve a confession to make. I was wrong. Last week I thought Claude Code was expensive. I burnt $5 in an hour. I compared it to the $18 monthly cost of Claude and concluded it was expensive. But it was the wrong comparison. I should have compared it to a software engineering salary. And that’s a very different comparison.
How so?
First we need a bit of context on software career progression in big tech. All the big tech companies have defined career progression paths. Engineers start their careers as a junior software engineer (SWE). Then, over the years they advance through the ranks to senior/ staff. Some folks make it to principal; a smaller set make it to distinguished engineers or technical fellows. And it’s lucrative. Even at entry level, the money is absurdly good. Microsoft pays ~$140k, Amazon ~$160k, Facebook & Google ~$180k.
Crudely, the principals design the architecture. The seniors break-up the architecture into components and tell the juniors what to write. The juniors then write the code.
But now we have Claude Code. The first of the agentic coders. And it’s competing with the junior roles. Let’s assume a junior works a 40 hour week for 50 weeks each year. That’s 2,000 hours. Of that they’re likely writing code 50% of the time. The rest of the time is meetings, training, education, illness. Maybe even some vacation.
1,000 hours per year writing code. From the employers point of view, $150k is equivalent to 1,000 hours of Developer Coding Time. Let’s call that DCT for short.
At $5 per hour that same $150k gets you 30,000 hours of Claude Code. That’s 30x more DCT. Thirty. Times. More.
But it gets worse. DCT isn’t the measure we care about. It’s lines of code we care about. A good human writes 100 lines of production ready code per day. Or 12.5 lines per hour. Claude Code wrote 1kloc in that hour. Claude Code is 80x more productive per unit time.
In total that makes Claude Code 2,400x more productive than a human. Three orders of magnitude more productive. Just ponder that for second. It’s insane.
Really? Are you sure?
You can pick holes in my calculation. Maybe the coding rate isn’t quite 1kloc per hour. Maybe Claude Code is a bit more expensive than $5 per hour. I agree. 2,400x is wrong. But it’s not three orders of magnitude wrong. Even if Claude Code was only 2x more productive than a human it’s still a game changer.
And I’ll get less wrong over time. The cost per 1M tokens has dropped by ~80% in the last twelve months. If that continues $150k will buy 6x more Claude Code in a year’s time. And look at the improvement in models over the past 12 months. That’s set to continue. It won’t be 1 kloc per hour. It’ll be two. Or four. Or more. You’ll have a choice of agentic coders. All the other big AI players will have their own versions by then. We’ve already got Gemini Coder.
Have no doubt. Software engineering is going to completely change. If I’m honest, deep down a part of me would like to be wrong. I don’t want my industry to change so radically. But it will. It has to. No one can ignore this level of productivity boost.
In the short term we’ll struggle to exploit the boost these new agentic coding tools provide. We don’t have enough seniors to feed Claude Code. But the solution to that will likely be Claude Senior Engineer. And then Claude Principal Engineer.
It seems inevitable the entry level SWE roles will cease to exist in a few years. What does that mean for folk currently in those roles? What does it mean for the folk at university developing those skills? As I wrote earlier this week we’re going to need to invest in significant re-training. In the short term developing the skills to drive the agentic coders will have significant value. But it won’t be long before those skills too are rendered unnecessary by AI.
How does this play out?
It’s always interesting to look at the limits when considering the future.
At one extreme, companies continue developing code at the same rate as present, but increasingly replace humans with AI.
At the other companies retrain employees to take advantage of AI and start producing y times more code (and by implication products) than previously. Where y could plausibly be 1000x greater.
The reality inevitably lies somewhere in the middle. But in both scenarios many software engineers become redundant. That much is clear.
And it’s not just software engineers who’ll be affected.
Products are going to be overhauled. Claude Code reimagines the coding workflow - it’s a “chat-coder”. But don’t imagine for a moment it’s just code that will change. We’ll get “chat-document-editors” and “chat-spreadsheets”. These will be new, re-imagined workflows - not AI bolted onto the side of legacy apps. Email is set to disappear, replaced by AI agents that talk to each other. The interface to calendars, to-do lists, task tracking, project management tools. They will all be re-imagined.
Companies are going to have an increasingly difficult time keeping up. Updating your AI policy is going to become a monthly and then weekly activity. You cannot risk falling behind.
The value of software will drop significantly. When it’s quicker and easier to build on demand, why buy in a 3rd party product? I’m already doing that for small tools; it will inevitably scale up over time.
And?
For the past six months I’ve been writing about the changes the software industry is going to face over the coming years. I’ve accepted change is coming. What I’m only just realizing is how fast the change is coming. How comprehensively things are going to change. Claude Code makes it clear this is a fundamental reinvention of the entire industry. The economics are too compelling to ignore.
For individuals, the path forward has to be: understand how to use AI effectively, learn system design, learn AI orchestration, keep up. Learn to be the conductor rather than just another musician. Cultivate the uniquely human abilities that remain valuable—empathy, creativity, ethical judgment, and contextual understanding.
For organizations, the challenge is even greater. How do you restructure your entire development process around these new capabilities? How do you retrain your workforce? How do you ensure your products remain relevant? How do you maintain quality and security when code generation becomes nearly free? The companies that solve these problems first will gain significant advantages.
If you work in software, don’t kid yourself this won’t happen. If you’re not fully engaged, don’t fall into the trap of thinking there’s plenty of time. The past six months have shown the rate of progress is accelerating and shows no sign of slowing down. And, remember, even if AI progress stopped today, the tools available right now are still so powerful they will upend the industry.
For all of us who work in software, change is coming, whether we like it or not. It’s exciting. And, if I’m honest, quite unsettling. But, to survive, we’re all going to need to find ways to remain relevant in this new world.