PilotDeck 2.0: Open-Source AI Agent OS with Inspectable Memory and 83% Cost Reduction

Tsinghua University's THUNLP lab, alongside ModelBest and OpenBMB, has released PilotDeck 2.0, an open-source operating system for autonomous AI agents distributed under the AGPL-3.0 license. The system directly tackles context contamination by strictly isolating project files, skills, and memory inside dedicated WorkSpaces, while introducing hierarchical multi-model routing that slashes execution costs on complex workloads by up to 83% compared to monolithic flagship models.

Isolated WorkSpaces Instead of Context Contamination

Anyone who runs autonomous agents across multiple client repos knows how fast context pollution ruins output. Ask an assistant to patch a database schema, then switch to client copywriting in the same session, and instructions inevitably cross wires. PilotDeck solves this by building around dedicated WorkSpaces. Each project operates in a sandboxed space with an independent file system, private memory database, and specific toolsets, ensuring that switching between concurrent tasks never corrupts historical context.

White-Box Memory and Background Dream Mode

Standard chat memory often feels like a frustrating black box: you cannot inspect what facts the agent extracted or where hallucinations crept in. PilotDeck surfaces memory records directly to the operator. Users can review stored preferences, edit misremembered constraints, and pin critical instructions without wasting time reprompting.

When the agent goes idle, a background daemon called Dream mode activates. It consolidates scattered turn-by-turn logs, removes duplicates, and organizes key milestones into clean summaries. If automatic compaction drops an important detail, users can instantly roll back to the pre-consolidation checkpoint.

Leader-Worker Orchestration Cuts Inference Costs

High inference bills remain the biggest hurdle for long-running agent loops. Firing a monolithic flagship model for every trivial step—from high-level planning down to text cleanup—drains budgets in hours. PilotDeck implements dynamic smart routing: top-tier models handle strategy and task decomposition, while lightweight models manage file edits, data formatting, and drafting.

Across seven benchmark workloads (including multilingual podcast releases, multi-source financial synthesis, and codebase architecture mapping), a single flagship model scored 69.1 points at a steep cost of $18.36 per run. PilotDeck's leader-worker hierarchy achieved 70.6 points while lowering the bill to $3.15. On the Workbuddy benchmark, PilotDeck running with Qwen3.6-35B-A3B reached 79.1 points, outscoring Claude Code at 77.3.

Platform metrics from our production systems over the past 30 days (September 18 to October 18, 2026, sampling 18,156 token transactions and 12,559 completed jobs, plus 600 active code projects over 90 days) confirm that overspending stems almost entirely from using flagship reasoning models on mundane formatting and parsing steps.

Always-On Execution and Realistic Trade-Offs

The system does not require an active browser tab. In Always-on mode, agents pursue scheduled tasks in the background, identify missing subtasks autonomously, and save compiled files and executive summaries straight to disk. Native MCP support ensures clean interoperability with existing developer tools and CLI environments.

The trade-offs are worth noting: the AGPL-3.0 license demands open-source distribution for derivative systems, which limits proprietary closed-source adoption. Furthermore, for quick single-turn queries, orchestrating multiple subagents introduces noticeable latency compared to simple direct prompting.