Why Fully Agentic Development is Unsolvable

Principal-Agent Problems with Dynamic, Unobservable, and Fractal User Intent
Rex W. Douglass, PhD

For 39 days I ran my own home agent system, a head I talk to, a state machine for every piece of work, an optimizer across eight model providers, and a memory, and measured where delegating to it broke. User intent is fractal, latent and dynamic, and an agent's work is unverifiable by the person delegating to it. Agents can shrink the human's part, never remove it.

Some findings

1. For requests under 25 words, the agents wrote about 11 words of “what I meant” for every word I wrote (866 requests).
2. Two models writing down what “done” means for the same 434 requests agreed on only 53% of it.
3. The system's own check said 3 of 3 worksheets were recorded; the real tablet had recorded none of 28.