Dev builds open ad-free site running 38 controlled experiments on AI agent patterns
meabs72 · reddit · 2026-11-06
A developer built an open, ad-free site with 38 controlled experiments testing popular agent harness patterns under real constraints: cheap models, limited tokens. Experiments cover planner/coder/reviewer crews vs single agents, parallel coding agents, gameable security triage, and whether a generated backend guide helps small local models. He's soliciting honest feedback on experiment design, cost/failure documentation, and which "obvious" agent patterns deserve proper trials.
More from coding & agent
- OpenClaw Ecosystem Hits 47 Products; Users Say Agents Outgrew the Chatbot Label — heyneighbor · 2026-11-06
- Full workflow: idea to finished AI video with Claude Code/Codex + Seedance, minimal re-rolls — dotey · 2026-11-06
- New vibe coder replicates Autodesk Maya in 3 hours with ChatGPT, asks for better workflow — Kiiaro · 2026-11-06
- Anthropic engineer confirms he's officially building Claude Code Desktop — maxleiter · 2026-11-06
- "Did your agents patch today's CVE?": dev shows agent running in the dark — aniketmaurya · 2026-11-06
- Dev fixes trivial bugs just to stop Claude nagging about them — jmugan · 2026-11-06