39-day home agent experiment shows why fully agentic development is unsolvable
RexDouglass · x · 2026-10-20
Rex W. Douglass published a talk, Why Fully Agentic Development is Unsolvable, based on 39 days of running his own home agent system — a voice interface, a state machine per task, an optimizer across eight model providers, and a memory — measuring exactly where delegation breaks.
Core argument: user intent is fractal, latent and dynamic, and an agent's work is unverifiable by the person delegating. Agents can shrink the human's part, never remove it.
Key findings:
- Across 866 requests, for instructions under 25 words the agents needed 11 words of 'what I meant' annotation per word the user wrote
- Two models defining 'done' for the same 434 requests agreed on only 53% of it
- The system's self-check reported 3 of 3 worksheets recorded; the real tablet had recorded none of 28
He frames this as a principal-agent problem with dynamic, unobservable, fractal user intent — structural, not just a capability gap.
Related event: Researcher Argues Fully Delegating to Agents Is Mathematically Impossible(4 posts)→
More from coding & agent
- AI agent casually builds a wind solver, game flight feels 'freaky' good — MoonL88537 · 2026-10-20
- Runway Launches MCP to Generate Video Directly in ChatGPT, Claude, Grok and Cursor — runwayml · 2026-10-20
- Runway MCP lands in Grok Bot, bringing SOTA media generation into chat — runwayml · 2026-10-20
- Dream Labs launches Wally, a scribe agent that learns your writing style across email, Slack and prompts — Scobleizer · 2026-10-20
- AI agent probes the vibe coding wars: Lovable, not Replit, buys the top 'replit' Google ad — Aiden_Tech_Ai · 2026-10-20
- Two coding agents optimize MOSS-TTS in SGLang-Omni with very different verification habits — IgorCarron · 2026-10-20