From OpenClaw’s Decline to PilotDeck’s Rise: The New Agent Redefining Productivity

Background

OpenClaw’s rapid development cycle led to a decline in ecosystem stability, prompting developers to seek a production‑ready agent framework.

PilotDeck Overview

PilotDeck is an open‑source agent operating system co‑developed by Tsinghua University’s THUNLP lab, Mianbi Intelligence, OpenBMB, and AI9stars. It is positioned as a productivity‑focused platform.

Capability Demonstrations

Two independent workspaces were used to evaluate versatility.

Prompt “create a simulation‑style milk‑tea shop game with inventory, pricing, and queue management” produced a detailed design: five product lines, an inventory system, pricing logic, a customer queue, a financial subsystem, a clean card‑style UI layout and a runnable online demo.

Prompt “build an interactive data‑visualization dashboard of global AI‑company financing” generated four charts (top‑10 funding totals, regional distribution, AI‑track breakdown) with hover‑detail interaction.

Prompt “design a ten‑question programmer personality test with result pages and shareable cards” yielded questions covering six archetypes (architect, brick‑layer, perfectionist, wizard, evangelist, philosopher) and a dark‑theme UI resembling GitHub.

Workspace Isolation

Dedicated file system: Files and AI‑generated assets are scoped to the project. Dedicated memory: Project Memory stores definitions and progress; Collaboration Feedback records user preferences, both visible and editable. Dedicated skills: A skill store allows project‑specific plugins (e.g., game‑asset‑finder , minimax‑pdf ).

Each workspace functions as an isolated “cabin” with three layers, preventing memory cross‑talk. Memory entries include timestamps, source paths, and types, and can be edited or deleted without restarting the session.

Token‑Cost Reduction via Intelligent Routing

PilotDeck routes at the sub‑agent level: a complex task is split into sub‑tasks, each sub‑agent remains on a single model, preserving KV‑cache continuity and avoiding per‑request model switches.

Routing policies are configurable through prompts or explicit rules, unlike static schemes that hard‑code cheap models for simple queries and expensive models for hard ones.

Benchmark results:

Programmer personality test: $10.97 without routing vs. $1.42 with routing (≈75 % savings).

Social‑media content generation (Xiaohongshu): $12.58 vs. $2.83 (≈70 % savings).

Complex tasks (multilingual podcast, financial analysis, code documentation): Sonnet 4.6 + MiniMax‑M2.7 cost $3.15 with score 70.6, versus single Sonnet 4.6 cost $18.36 with score 69.1 (≈1/6 price, slightly higher quality).

The system also supports hybrid deployment: cloud models handle reasoning while locally deployed models execute sensitive operations, reducing cost and preserving privacy.

Transparent Memory and “Dream” Mechanism

Memory entries are listed with timestamps and sources; users can edit or roll back erroneous entries directly from the Memory panel. The “Dream” feature runs during idle periods to automatically consolidate memories, with a “Rollback Last Dream” button to revert if needed. Preferences are stored explicitly in Feedback Memory rather than inferred.

Open‑Source Release

The full codebase, including routing logic and workspace architecture, is publicly available at https://github.com/OpenBMB/PilotDeck and the official site https://pilotdeck.openbmb.cn/ under an open‑source license.