Blog
March 12, 2026Philosophy6 min

Vibe Coding With a System

1,087× productivity, 0 lines of manual code — an empirical analysis of the agent-in-the-loop paradigm

Claude Code · MCP · Agentic Coding

I don't write code anymore. Not because I can't. But because in the time it takes me to type a function, I can have three agents working on three features in parallel. My role has fundamentally changed. I'm no longer a programmer. I'm a conductor.

The Study Nobody Asked For

In early 2026, I started analyzing my own workflow. Not as a thought experiment — empirically. Every interaction with AI agents systematically evaluated: What creates value? What's overhead? Where do I lose time?

Raw Data

  • 2,600 messages
  • 55 sessions
  • 671 MB transcripts
  • 6-month period

Results

  • 38% frustration overhead
  • 31% automatable
  • 6% genuine strategic decisions
  • 23.8% bug-fix resolution rate
38%of my communication with agents is overheadRepetitions, clarifications, re-providing context that should already be known.

Human in the Loop Is Wrong

The established paradigm is called "Human in the Loop" — the human as a control instance in an AI-driven process. My data shows: that's the wrong perspective.

The human is not the control instance. The human is the primary actor. The AI agents are interchangeable tools in their loop.

Human in the Loop

  • AI drives the process
  • Human controls
  • Human as safety net
  • Reactive

Agent in the Loop

  • Human drives the process
  • Agents execute
  • Human as strategist
  • Proactive

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 is no music.

1,087 Developers in One Head

1,087×Productivity multiplierOne person with AI agents produces the output of 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 — quality isn't 1,087 times better.

Agents make mistakes. They hallucinate. They lose context. 21% of my messages concern lost context. Bug-fix loops have a resolution rate of only 23.8%. UI bugs only 20%.

The Seven Roles of the Conductor

From the analysis, I extracted seven core functions that the human assumes in AI-assisted development:

Vision

Where are we going? What are we building and why?

Quality

Is this good enough? What's missing? What's wrong?

Context

What doesn't the agent know that it needs to?

Architecture

How do the pieces fit? What breaks otherwise?

Risk

What can go wrong? What does a mistake cost here?

Prioritization

What matters now? What can wait?

Trust

When to trust the agent — and when not to?

None of this is programming. All of it is leadership.

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 anymore. But that's exactly the point: I don't drive anymore — I navigate. And the vehicle is faster than anything a human could steer alone.

The role of the 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 orchestra music has no business hours.

— Philipp

Want to understand how AI agents can transform your development?

Book a call