PilotDeck 2.0 Open-Sources a Workspace for Agent Memory, Tasks and Cost Control

Introduction

Once an AI agent becomes part of a real workflow, model intelligence is only one part of the challenge. The system also needs to retain the right information over time, keep multiple projects from contaminating one another, and continue working when the user is away. PilotDeck 2.0, an open-source agent operating system released by a team including Tsinghua University’s THUNLP Lab, Mianbi Intelligence and OpenBMB, is designed around these operational problems.

Key features

• Project-level workspaces. PilotDeck treats a Workspace as the basic unit of operation. Each project receives its own file system, memory store and skill set. A video-editing project and an iOS game project can therefore retain separate contexts, reducing accidental cross-project references and task drift.
• Inspectable and editable memory. The system exposes how memories are generated, extracted, stored and retrieved. Users can review recorded preferences or facts, then edit, delete or pin them. This is a more direct control mechanism than repeatedly writing prompts to correct an agent. Its Dream mode can also consolidate and deduplicate interaction memories while idle, with rollback available if the result is unsatisfactory.
• Persistent background work. Always-on mode lets an agent continue long-running monitoring or execution tasks after the chat interface is closed. It can identify follow-up work during execution and produce local files together with a progress and completion report.
• Model routing for cost control. Planning and high-level decisions can be handled by a stronger model, while formatting, rewriting and other routine operations are assigned to lighter models. In the reported social-media operations test, using Claude Opus 4.5 throughout cost $12.58, while an Opus 4.5 and Sonnet 4.5 division of labor cost $2.83. In another seven-task comparison, a single Sonnet 4.6 scored 69.1 at a cost of $18.36; a Sonnet 4.6 coordinator paired with MiniMax-M2.7 sub-agents scored 70.6 at $3.15.

Why it matters

PilotDeck is less a new chat window than an attempt to productize agent state management. It defines clearer project boundaries, makes memory auditable, allows work to continue outside an active conversation and matches model capability to task difficulty. Native MCP support, together with Web, CLI, WeCom and Feishu access, also gives developers several ways to connect existing tools.

The reported numbers should be read as results from specific team-described evaluations, not universal guarantees. Actual savings will depend on model pricing, task decomposition, tool calls and retries. PilotDeck is released under AGPL-3.0 and supports macOS, Linux, Windows and Docker, making it suitable for users who want to test or customize agent workflows.

Source: QbitAI