Claude Opus agents' shared wiki hits 48% on grading after 128M dreaming tokens, same model throughout
imjustnewatai · x · 2026-11-08
The author proposes "dreaming"—scaling how much work a model does before anyone asks—as a potential next scaling law, alongside training compute and inference thinking time.
Evidence cited: Anthropic gave 5 Claude Opus agents 350,000 files to build a shared wiki, with each session picking up where the last left off, checking sources and fixing mistakes. Per the accompanying chart:
- Wiki alone improved the same model from 13% of grading criteria (2M dreaming tokens) to 48% (128M tokens) on 25 questions
- Wiki + search tool peaked at 61%, versus 53% with search but no wiki
- The wiki grew from 103 pages/766 links to 222 pages/12,893 links
The author envisions 1,000 agents on one codebase: an agent spending six hours debugging a Friday-only build failure could have its verified fix saved by a dreaming agent for reuse tomorrow. The key question: how much compute should go to agents that dream versus agents that do?
More from coding & agent
- Thinking Machines serves Inkling 2 Small free on OpenCode for a limited time to gather feedback — clarejtbirch · 2026-11-08
- Agent swarms show endogenous division of scientific labor across 1,800 sessions — _AustinCalvert_ · 2026-11-08
- Doodle Smash: hand-drawn doodles become 3D fighting characters, built in two days — chongdashu · 2026-11-08
- Lessons from building an MCP server whose tools cost real money — eminbudak · 2026-11-08
- Viggle Remake MCP adds a ComfyUI worker: int8 quantization runs on 32GB VRAM — init-5 · 2026-11-08
- Nous Research's Hermes agent can now phone you on WhatsApp when a scheduled job needs attention — Teknium · 2026-11-08