Writing on Dependable Agents: an experiment from 100% AI down to 51%
BabyKai26 · reddit · 2026-10-17
A Redditor wrote an article on Dependable Agents as an AI-assisted writing experiment: the first draft was 100% AI-generated, and after manual revisions it tested at 51% AI / 49% human per Pangram. Key observation: AI struggles to break out of trained writing patterns, so the author reviewed manually while saving lessons into a skill for reuse. Link in comments.
More from coding & agent
- Sooool: coding agents that reason about real-time API latency — wzenus · 2026-10-17
- MVP talk at AI Engineer: what does the system do, and can you change it safely? — altryne · 2026-10-17
- User makes AI assistant file an FTC complaint over failed $1M pizza giveaway — arthurcolle · 2026-10-17
- Fiddler's Dynamic Model Escalation: cheap judge scores everything, reasoning model sees only deferrals — hima_lakkaraju · 2026-10-17
- Design agent approvals around risk, not org charts — usehive · 2026-10-17
- Stealth Browser MCP (2.2k stars): AI Agents Bypass Cloudflare via Real Chrome — tom_doerr · 2026-10-17