Blind test of 5 self-hosted AI agents on Qwen 3.8 27B: one fails by asking permission
GapNew4766 · reddit · 2026-10-26
A Reddit user blind-tested 5 self-hosted agents (OpenCode, goose, Atomic Agent, Hermes, OpenClaw) on an M5 Max with 64GB, all running Qwen 3.8 27B in llama.cpp, across 5 tasks: contact merging, function renaming, bug fixing, log error counting, and receipt-to-spreadsheet. Four agents passed everything; one failed by stopping to ask 'can I install this?' with no one watching. Fastest total: 10:29. The author asks: would you rather have an agent that installs autonomously or asks first?
More from coding & agent
- Epic's AI MegaJam offers $65K prizes for AI-agent-built games, Nov 12-15 — lmoroney · 2026-10-26
- A garage door alert shows how proactive AI agents drown your inbox — manosaie · 2026-10-26
- Full voice agent (STT+LLM+TTS) on one A40: 1.28s median latency, tuning details shared — Broad_Breakfast_1172 · 2026-10-26
- One prompt in, full music video out: Claude autonomously drives ComfyUI pipeline — Cheap_Credit_3957 · 2026-10-26
- t3 code's clean API lets you hack the harness without recompiling the app — pvncher · 2026-10-26
- Dev vibe-codes parents' TV ad-free with Claude Code, hits 34,800 GitHub stars in 8 days — 141_1337 · 2026-10-26