dottxt's ^Ctrl: zero rule violations in 3,508 agent calls vs 41 for the baseline
remilouf · x · 2026-10-22
dottxt showed that on Sierra's τ²-bench, a Qwen3.8-27B agent given the rulebook broke it 41 times, but with the ^Ctrl framework: zero violations across 3,508 calls, with more tasks solved on every attempt.
More from coding & agent
- Karpathy open-sources AutoResearch: AI agents that run the full research loop autonomously — mdancho84 · 2026-10-22
- Dev uses Claude Opus 5.5 to build a working replica of ILM's Stagecraft volume in Three.js — jasonkneen · 2026-10-22
- Solo founder ships one-command tool to find X's biggest Claude Code and Codex token burners — ThePeterMick · 2026-10-22
- Open-Source Claude Recruiter Finds Ranked Candidates with Personal Emails at $0.20 Each — brandon_galang · 2026-10-22
- Retention, not loops: why recursive self-improvement keeps starting over — BandAutomatic2009 · 2026-10-22
- cinetic: open-source agent skill turns a one-line brief into a branded launch film — aigclink · 2026-10-22