How PilotDeck’s Open‑Source Agent Cuts Token Costs by 70% with Parallel Workspaces

Overview

PilotDeck is an open‑source intelligent‑agent operating system jointly developed by Tsinghua University’s THUNLP Lab, Mianbi Intelligent, OpenBMB and AI9stars. It was released to address the high token costs and poor isolation of earlier agents such as OpenClaw.

WorkSpace Architecture

PilotDeck introduces a WorkSpace concept. Each WorkSpace is a self‑contained “agent cabin” with three independent layers:

Dedicated file system – files generated by the agent are stored in a private directory.

Dedicated memory – Project Memory records project definitions and progress; Collaboration Feedback records user preferences. All entries are visible with timestamps, source paths and types, and can be edited or deleted.

Dedicated skills – a skill store allows installing skill packages (e.g., game-asset-finder, minimax-pdf) into a specific WorkSpace.

Demos

Prompt: “make a simulation game of a milk‑tea shop with inventory, pricing, queue logic, UI layout and JavaScript modules.” PilotDeck generated a complete design: five product lines, pricing rules, a queue system, a financial subsystem, a clean card‑style UI, and the key JS modules needed to run the game online.

Prompt: “create an interactive data‑visualisation screen for global AI‑company financing with animation and hover details.” PilotDeck produced four charts – TOP‑10 funding, regional financing shares (North America, Europe, Asia), and distribution across AI tracks (general, enterprise, generative) – and assembled them into a coherent dashboard.

Prompt: “design a programmer personality test with 10 questions, classify users into six personas, and render the result with a GitHub‑dark theme and JetBrains Mono font.” PilotDeck generated ten realistic multiple‑choice questions, mapped answers to six personas (architect, brick‑layer, perfectionist, wizard, evangelist, philosopher), and created a dark‑theme result page ready for sharing.

All three WorkSpaces run concurrently without interfering with each other.

Memory System

The memory panel lists each memory entry with a timestamp, source path and type. Users can edit or delete incorrect memories directly. A “Dream” mechanism automatically consolidates memories during idle periods, and a “Rollback Last Dream” button restores the state before consolidation.

Smart Routing

PilotDeck’s routing operates at the sub‑agent level rather than per request, preserving KV‑cache continuity for each sub‑agent. Routing rules can be defined with natural‑language prompts (e.g., “code‑related sub‑tasks use Claude Opus, text tasks use the cheap model”).

Programmer‑personality test: cost reduced from $10.97 to $1.42 (≈75% saving).

Social‑media content generation (Xiaohongshu): cost reduced from $12.58 to $2.83 (≈70% saving).

Complex tasks (multilingual podcast, financial analysis, code documentation): combined Sonnet 4.6 + MiniMax‑M2.7 cost $3.15 with a score of 70.6, versus a single Sonnet 4.6 cost $18.36 with a score of 69.1 – about one‑sixth the price with slightly better quality.

Routing decisions are displayed in a Routing panel, showing task difficulty (simple/medium/complex), estimated cost without routing, and actual cost after routing.

Local Model Integration and Tool Provisioning

PilotDeck can attach locally deployed models as sub‑agents, keeping privacy‑sensitive data on‑premise. It can also provision tools automatically; for example, when processing a multilingual podcast, PilotDeck installs the VoxCPM voice‑synthesis model on the client side while the cloud model handles reasoning.

Comparison with Other Agents

Claude Cowork’s “Projects” and Cursor’s “Workspace” isolate only folders and static rules. Their memories are opaque, skills do not evolve with usage, and cost attribution per project is unclear. In contrast, PilotDeck’s WorkSpace isolates an entire execution environment (file system, memory, skills) and makes all memory entries visible and editable.

Open‑source Availability

All code, including routing logic and WorkSpace architecture, is publicly available at:

https://github.com/OpenBMB/PilotDeck

Official site (documentation and releases):

https://pilotdeck.openbmb.cn/