The THUNLP Laboratory at Tsinghua University, ModelBest, OpenBMB, and AI9Stars have introduced PilotDeck 2.0—a new version of their open-source agent operating system built on isolated project workspaces. The update was published on October 15th and received coverage in the Chinese AI publication QbitAI. The code is available on GitHub under the AGPL-3.0 license and functions on macOS, Linux, and Windows operating systems, while also supporting deployment via Docker.
According to the project's README file, version 2.0 includes one-click installation and visual configuration. The system is capable of supporting long-term tasks for up to 2000 interaction messages and comes with new skills for creating files in Word, Excel, and PowerPoint formats.
Architecture and Context Management
The main structural element is the Workspace. Each project is allocated its own set of file system, memory storage, and skill set. This allows an agent editing video in one workspace and an agent creating a mini-game for iOS in another to completely separate their contexts. Developers position this as a solution to the problem of context contamination, where instructions from one task subtly influence the next.
Memory is designed as a white box, giving users the ability to view every saved agent entry, the time it was created, and the workspace it belongs to. Users can directly edit, delete, or pin entries instead of writing prompts that require the agent to forget information. The Dream Mode feature consolidates and eliminates duplicates of scattered memories while the agent is idle, offering a one-click rollback option in case the cleaning process leads to the loss of important details.
The Smart Routing system enables strong model planning by delegating formatting, polishing, and other routine steps to lighter models. In a social media scenario, the team reported that using Claude Opus 4.5 as the primary agent and Sonnet 4.5 as a sub-agent reduced the cost of a single run from $12.58 to $2.83. When performing seven more complex tasks, including creating multilingual podcasts, multi-source data reports, and codebase documentation, the combination of Claude Sonnet 4.6 and MiniMax-M2.7 achieved a score of 70.6 at a cost of $3.15, compared to 69.1 at a cost of $18.36 for Sonnet 4.6 alone, representing savings of about 83%.
On the WorkBuddy Bench PilotDeck, utilizing Qwen3.6-35B-A3B, scored 79.1 points, while Claude Code scored 77.3. All presented figures are self-reported by the team.
The Always-on mode allows agents to continue working in the background after the user closes the chat, performing long-term monitoring tasks, finding potential assignments, and saving results to local files with a brief progress overview. The system natively supports the Model Context Protocol and behaves consistently in both web interfaces and collaboration applications such as WeCom and Feishu.