Anthropic’s Own Data Proves Harness Theory — Humans Frame, AI Executes

Anthropic just published research analyzing 400,000 Claude Code sessions. The finding that matters most: humans make 70% of planning decisions. Claude handles 80% of execution. Anthropic’s own data just validated the exact division of labor that Harness Theory describes.

400,000 Sessions — The Division of Labor

The Human

70%

of planning decisions

“what to do”

Claude

80%

of execution decisions

“how to do it”

This Is Harness Theory, Measured

In Life in the Harness, we wrote:

Life in the Harness — Published June 13

“The frame is the whole of the human contribution. Everything downstream of it is execution.”

Anthropic’s data says the same thing with 400,000 data points: 70% frame, 80% execution. The human decides what to build. The agent decides how to build it. This isn’t theory anymore — it’s the measured reality of how humans and AI work together.

The Expertise Gap Is Real

Output Per Prompt — By Expertise

Novice~5 actions, 600 words
Expert~12 actions, 3,200 words

Experts get 5x more output from the same AI. The frame determines the yield.

Experts trigger 2-3x more actions per instruction and get 5x more output than novices — from the same model. The AI is identical. The frame is different. This is the Framing Ladder in data: the edge isn’t the tool, it’s the frame you bring to it.

Critical finding: Domain expertise matters more than coding background. Software engineers achieve 34% verified success; non-software occupations reach 29%. All major occupations cluster within 7 percentage points. The AI levels the technical playing field. What’s left is judgment.

What’s Changing

The data also shows how work itself is shifting (Oct 2025 → Apr 2026):

Debugging33% → 19% ↓
Operating software14% → 21% ↑
Data analysis + writing10% → 20% ↑ (2x)
Average task value+27% overall, +43% for building tasks

Debugging is collapsing. Building and analysis are expanding. The work is moving up the value stack — less fixing, more creating. This is the Framing Ladder in motion: as the AI handles more execution, humans move to higher-value framing work.

Business Engineer

My Life in the Harness — The Framework Behind the Data

The Framing Ladder, the seven harness principles, and the complete system for the agent era — now validated by 400,000 sessions of real-world data.

Read: Life in the Harness →

The Bottom Line

Anthropic studied 400,000 sessions and found that humans frame, AI executes. Experts get 5x more from the same model. Domain knowledge beats coding skill. Debugging is dying; building is growing. Every data point validates what we’ve been writing: the model is commoditized. The frame is the moat. The harness is the system. And expertise — real expertise — has never been more valuable.

Source: Anthropic Research — Claude Code & Expertise

Scroll to Top

Discover more from FourWeekMBA

Subscribe now to keep reading and get access to the full archive.

Continue reading

FourWeekMBA