Tsinghua's THUNLP, ModelBest and OpenBMB release open-source PilotDeck 2.0 agent OS
Tsinghua's THUNLP lab, together with ModelBest, OpenBMB and other teams, released PilotDeck 2.0, an open-source agent operating system built around WorkSpace units that isolate each project's file system, memory store and skill set, according to QbitAI. The system makes memory "white-box" — users can view, edit, delete and pin entries — and adds a Dream mode that consolidates scattered memories while idle plus an Always-on mode that keeps long-running tasks going after the chat window is closed; it natively supports MCP across web, CLI and IM entry points such as WeCom and Feishu. The team says PilotDeck scored 79.1 on Workbuddy Bench with Qwen3.6-35B-A3B versus 77.3 for Claude Code, and that on seven hard long-horizon tasks a "Sonnet 4.6 planner plus MiniMax-M2.7 worker" routing setup scored 70.6 at $3.15, against 69.1 at $18.36 for standalone Sonnet 4.6. PilotDeck is released under AGPL-3.0 for macOS, Linux and Windows, with a Docker deployment option.
The team's own comparisons indicate that routing complex planning to a strong model and execution to a cheaper one can cut cost to roughly one-sixth while slightly exceeding a single top-tier model on its test tasks — a design the project ships inside an open-source, MCP-native agent operating system.
🌊 The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
An agent meta-harness for Claude Code and Codex.
Get started with Ruflo
#PilotDeck#Tsinghua THUNLP#OpenBMB#ModelBest#Qwen3.6-35B-A3B#Claude Code#Claude Sonnet 4.6#Claude Opus 4.5#MiniMax-M2.7#Model Context Protocol (MCP)#Ruflo