THUNLP Open-Sources PilotDeck 2.0: Workspace Isolation + Smart Routing, Agent Task Cost to 1/6?

TL;DR

Tsinghua University's THUNLP Lab, together with ModelBest (面壁智能) and OpenBMB, open-sourced the agent operating system PilotDeck 2.0 on October 17, 2026, under AGPL-3.0 license, supporting macOS, Linux, Windows, and Docker deployment. It targets the three most painful problems when using AI agents: files and memory getting tangled across projects, tasks stalling when humans step away, and token bills skyrocketing as tasks get complex. Four core features — WorkSpace isolation, white-box memory, Always-on mode, and Smart Routing — reduce complex task costs to approximately 1/6 of a top-tier single model.

Four Core Features

1. WorkSpace: From "Cross-Contamination" to "Isolation

Most agent tools have black-box memory: users don't know what it remembered, when, or where. When projects multiply, Project A's preferences contaminate Project B's context — you just finished writing a technical architecture, then ask for marketing copy, and the output still carries "architecture decoupling."

PilotDeck's solution makes workspaces into independent workstations:

• Each workspace has its own file system
• Each workspace has its own memory store
• Each workspace has its own skill set
• Memory and retrieval spaces are physically isolated
• Example: audio/video editing and iOS mini-games can run simultaneously without interference

This design isn't flashy, but directly addresses the multi-project switching needs of real work.

2. White-Box Memory: Viewable, Editable, Deletable

More critically, memory is white-boxed. Users can:

• View each remembered preference and fact item by item in a panel
• Directly modify or delete incorrect memories
• Pin important information with one click
• Dream mode: automatically summarizes and deduplicates scattered memories during idle time
• Supports one-click rollback

Compared to repeatedly writing prompts telling it to "forget wrong information," this editable memory is closer to human work notes.

3. Always-on Mode: Keep Working After Window Close

Always-on mode lets the agent continue executing long-cycle tasks after the window is closed. This is a key capability for agents moving from "demo" to "production tool" — real tasks often require tens of minutes or even hours of execution time. If the agent stalls the moment a human steps away, it can only ever be used in demos.

4. Smart Routing: Strong Model Decides, Small Model Executes

The biggest deterrent for long-running agent tasks is often the bill. Many tasks aren't difficult — just formatting, translation, or file processing — but simplistic system architecture calls the most expensive flagship model throughout.

PilotDeck has built-in task difficulty recognition, routing different stages to different-tier models:

• Main agent handles top-level planning → uses flagship model
• Sub-agents handle text polishing etc. → downgrades to cheaper model
• Automatically identifies task difficulty and routes accordingly

Cost Comparison Benchmarks

Social Media Operations Scenario

Seven Complex Tasks Comparison

Key finding: PilotDeck orchestration scored 70.6, slightly better than the pure top-tier single model's 69.1, at approximately 1/6 the cost ($3.15 vs $18.36).

Workbuddy Bench Evaluation

Using a smaller model, through orchestration-layer optimization, the score actually exceeded Claude Code — demonstrating that orchestration-layer optimization can deliver significant performance and cost improvements without upgrading to a larger model.

How to Use: Developer Action Guide

1. Deployment

AGPL-3.0 has certain restrictions on commercial use — if you plan to offer PilotDeck as a SaaS service, pay attention to license terms. But internal use and trials have no barriers.

2. Multi-Project Parallel Scenarios

If you manage multiple projects simultaneously (e.g., technical architecture and marketing copy), WorkSpace is your first line of defense. Create separate workspaces for each project, physically isolating files, memory, and skill sets.

3. Long-Cycle Task Scenarios

If your tasks require tens of minutes or hours of execution, enable Always-on mode. The agent continues executing after window close, and you check results when you return.

4. Cost-Sensitive Scenarios

If your agent bills are skyrocketing, enable Smart Routing. Benchmark data shows that while maintaining or slightly improving quality, costs can be reduced to approximately 1/6 of a single model. The key is configuring the main agent to use a flagship model and sub-agents to use cheaper models.

5. Memory Management Scenarios

If your agent's memory shows "cross-contamination" or "remembering wrong," use the white-box memory panel to view, modify, or delete items one by one. Dream mode can automatically summarize and deduplicate during idle time, reducing manual maintenance burden.

Industry Significance

From Model Capability to OS Capability

Agent competition is shifting from "model capability" to "operating system capability." Whoever can manage memory, isolate context, schedule models, and ensure continuous task execution will bring agents into production environments.

PilotDeck's four core features — WorkSpace, white-box memory, Always-on, and Smart Routing — are not model-level innovations but operating-system-level engineering optimizations. Yet this is precisely the capability most lacking for current agent deployment.

Another Data Point for Cost Curves

Published the same day as DeepSeek V4.1-Flash-1017 and GLM-Legal, all three paths differ but point in the same direction:

• DeepSeek uses Harness optimization to compress agent task runtime and output tokens
• GLM-Legal uses specialization to cut legal task costs to 1/70 of Gemini 4 Argon
• PilotDeck uses routing to split complex tasks across model tiers, cutting costs to 1/6

Competition is shifting from "who's smarter" to "who can keep agents working without going bankrupt."

Open Source Is Just the Starting Line

PilotDeck uses AGPL-3.0, lowering the barrier for enterprise trials. But open source is just the starting line — whether workspace isolation is thorough enough, whether white-box memory can scale, and whether routing strategies will misjudge in real business scenarios all need more practical validation. The war at the agent operating system level has only just begun.