The Conductor
Why I don't write code anymore — and still build more software than ever
I don't write code anymore. Not because I can't. But because in the time it takes me to type one function, I can have three agents working on three features in parallel. My role has fundamentally changed. I'm not a programmer anymore. I'm a conductor.
The study nobody asked for
In early 2026, I started analyzing my own workflow. Not as a thought experiment — but empirically. 2,600 messages, 55 sessions, 671 MB of raw data. Every interaction with AI agents systematically evaluated: What creates value? What's overhead? Where do I lose time? Where do I gain it?
The result surprised even me. 38% of my communication with agents is frustration overhead — repetitions, clarifications, providing context that should already be known. 31% could be automated. And only 6% of my interactions are real strategic decisions. The rest is conducting.
Human in the Loop is wrong
The established paradigm is called "Human in the Loop" — the human as a checkpoint in an AI-driven process. My data shows: that's the wrong perspective. The human isn't the checkpoint. The human is the primary actor. The AI agents are interchangeable tools in their loop.
I call it "Agent in the Loop." The conductor determines What and Why. The agents determine How. And like a real orchestra: the conductor doesn't play an instrument — but without them, there's no music.
1,087 developers in one head
The COCOMO analysis across all my projects shows an effort multiplier of 1,087. That's not exaggeration — that's math. One person with AI agents produces the output that would traditionally require 1,087 developer-months. 632,000 lines of code in OmniMindscape. 61,000 in FamilienKoch. In 9 days.
But — and this is the important part — the quality isn't 1,087 times better. Agents make mistakes. They hallucinate. They lose context. 21% of my messages involve lost context. Bug-fix loops have a resolution rate of only 23.8%. UI bugs even only 20%.
The seven roles of the conductor
From the analysis, I extracted seven core functions that humans fill in AI-assisted development: Vision and strategy. Quality control. Context management. Architecture decisions. Risk assessment. Prioritization. And perhaps the most important: deciding when to trust the agents — and when not to.
None of that is programming. All of it is leading.
What this means for the future
When I tell my consulting clients that I don't write code anymore, they look at me as if I'd said I can't drive a car anymore. But that's exactly the point: I'm not driving anymore — I'm navigating. And the vehicle is faster than anything a human could steer alone.
The role of software developer isn't disappearing. It's transforming. From craftsman to conductor. From code writer to systems thinker. Anyone who still believes a developer's value lies in their typing skills will have a problem in two years.
And me? I keep conducting. With five agents in parallel. At midnight. Because orchestral music has no business hours.
— Philipp