Why fully agentic development is unsolvable: four structural mechanisms, one talk
RexDouglass · x · 2026-10-20
In a new talk, Rex Douglass argues Anthropic's complaint that "people are bottlenecking agents" reflects an unavoidable structural problem, demonstrated via his agentic Jarvis and a homework-tablet case study: (1) user intent is fractal—every want splits into finer tradeoffs, like the coastline paradox; (2) intent is latent—you can't fully state it without building, and key qualities like "neat handwriting" have no direct measurement; (3) intent is dynamic—of 2,718 words that started intents in 110 messages, 2,738 words changed them after seeing results; (4) agent performance is unverifiable—his agent wrote 6,346 tests for its own code, 4,236 never failed and many were meaningless, while 27 days of build work implied 45 working days of reading (179,200 lines). Perfect delegation is a principal-agent dilemma, not a user skill gap.
Related event: Researcher Argues Fully Delegating to Agents Is Mathematically Impossible(4 posts)→
More from AGI Musings
- Gensyn co-founder to discuss auditable training, AI free of few-company control — benfielding · 2026-10-20
- He pitched pretraining-inference compute exchange rate research 9 months before o1 — RylanSchaeffer · 2026-10-20
- Author of 'Emergent Abilities a Mirage' says results valid but obvious to labs — RylanSchaeffer · 2026-10-20
- AI fluency boils down to reframing business bottlenecks as prediction problems — sanderssays · 2026-10-20
- Stewart Squared Ep. 111: When Robots Become Agents — StewartalsopIII · 2026-10-20
- AI rewards laziness over hard work — ask better questions, start from 80% — EXM7777 · 2026-10-20