Tsinghua-Led Team Open-Sources Agent OS PilotDeck 2.0, Cutting Complex Task Costs by ~5/6

Tsinghua University's THUNLP lab, together with ModelBest (Mianbi Intelligent) and OpenBMB, has open-sourced PilotDeck 2.0, an operating system designed for AI agents. The system uses "WorkSpace" as its core unit, providing isolated file systems, memory banks, and skill sets for different projects to prevent context pollution and memory crossover.

PilotDeck 2.0 features white-box memory management, allowing users to view the entire lifecycle of memory generation, storage, and retrieval, and to manually edit or delete entries. A "Dream mode" automatically organizes and deduplicates memories during idle time, with rollback support. The Always-on mode enables agents to continue long-running monitoring and execution tasks in the background after the chat window is closed, and generate completion reports.

The system natively supports the Model Context Protocol (MCP), keeping consistent backend behavior across web, CLI, and IM environments such as WeCom and Feishu, allowing developers to extend tools and workflows.

For cost efficiency, PilotDeck includes Smart Routing, which assigns complex planning and decisions to strong models while delegating routine execution and refinement to lightweight models. In tests on seven complex tasks, a "strong-main, weak-sub" configuration using Sonnet 4.6 as the main agent and MiniMax-M2.7 as sub-agents achieved a higher overall score than a single top-tier model, while costing about one-sixth as much. In the Workbuddy Bench evaluation, PilotDeck with Qwen3.6-35B-A3B scored 79.1, outperforming Claude Code's 77.3.

PilotDeck 2.0 is released under the AGPL-3.0 license and supports macOS, Linux, and Windows, with Docker deployment available.