CatalystTech/AI
Tsinghua open-sources! Agent operating system task costs reduced by 83%, AI memory searchable and editable
Tsinghua open-sourced an agent OS cutting task costs 83% with editable AI memory.
Context
What happened: Tsinghua's THUNLP lab, with ModelBest and OpenBMB, open-sourced PilotDeck 2.0, an agent operating system built around isolated WorkSpace units with white-box memory and smart model routing.
Key numbers: Workbuddy Bench score 79.1 with Qwen3.6-35B-A3B versus Claude Code at 77.3.
Why it matters: The project targets memory management, task execution, and model call costs, with routing that assigns complex decisions to strong models and simpler work to cheaper ones. It uses the AGPL-3.0 license and supports macOS, Linux, Windows, and Docker.
What changed
Tsinghua's THUNLP lab, with ModelBest and OpenBMB, open-sourced PilotDeck 2.0, an agent operating system built around isolated WorkSpace units with white-box memory and smart model routing.
Why it matters
The project targets memory management, task execution, and model call costs, with routing that assigns complex decisions to strong models and simpler work to cheaper ones. It uses the AGPL-3.0 license and supports macOS, Linux, Windows, and Docker.