Tsinghua's THUNLP open-sources PilotDeck 2.0, cutting complex task costs to about one-sixth

Tsinghua University's THUNLP lab has released PilotDeck 2.0 together with ModelBest, OpenBMB and other collaborators, publishing the agent operating system under the AGPL-3.0 licence. According to coverage by QbitAI, the project targets the three problems that bite hardest in day-to-day agent use: tangled memory, long tasks that stall, and runaway model costs.

The core abstraction is the WorkSpace. Each project gets its own file system, memory store and skill set, so an audio-editing workspace and an iOS game workspace never share retrieval context. That physical separation is meant to stop the context contamination that appears when several jobs run side by side.

Memory is treated as a white box. Generation, extraction, storage and retrieval are all visible to the user, who can open a panel and inspect exactly which preferences and facts were recorded. Mistakes can be edited, deleted or pinned directly instead of being argued away with a long corrective prompt.

Two further modes extend the system's reach. Dream mode consolidates, deduplicates and refines scattered interaction memories while the agent is idle, with one-click rollback if the tidy-up loses a detail. Always-on mode keeps long-running monitoring and execution going after the chat window is closed, then writes results to local files with a progress and summary report.

Cost control comes from Smart Routing, which classifies task difficulty and splits work between models. In a social-media operations test, running the whole workflow on Claude Opus 4.5 cost $12.58 per run; routing top-level planning to Opus 4.5 and sub-tasks to Sonnet 4.5 brought that down to $2.83, a saving of roughly 70 percent.

The gap widened on seven harder tasks spanning multilingual podcast distribution, multi-source reports, literature reviews and codebase architecture documents. A single MiniMax-M2.7 run was cheap at $1.90 but scored just 37.1; a single Claude Sonnet 4.6 run scored 69.1 at $18.36; PilotDeck's Sonnet-led, MiniMax-assisted arrangement scored 70.6 at $3.15, about one-sixth of the frontier-only bill and the basis for the headline claim of an 83 percent cost reduction.

On Workbuddy Bench, PilotDeck paired with Qwen3.6-35B-A3B scored 79.1, ahead of Claude Code at 77.3. The system supports MCP natively, and the same backend logic serves its web interface, command line and IM integrations such as WeCom and Feishu, so developers can extend existing workflows rather than rebuild them.

PilotDeck runs on macOS, Linux and Windows, ships with a Docker deployment option and lives in the OpenBMB organisation on GitHub. The benchmark numbers matter less than the architecture they demonstrate: memory, execution and model routing are being treated as operating-system problems rather than prompt-engineering ones. What to watch next is whether outside teams can reproduce these figures, and whether white-box memory holds up once real groups, not single users, start sharing a workspace.