---
id: "lib-2026-10-17-tsinghua-open-sources-agent-os-83-lower-task-cost"
type: "library"
title: "Tsinghua open-sources agent OS, 83% lower task cost"
title_original: "清华开源！智能体操作系统任务成本降83%，AI记忆可查可改"
lang: "ZH"
org: "量子位 QbitAI"
source_type: "Practitioner post"
date: "2026-10-17"
week: "2026-W42"
url: "https://nosov.ai/library/lib-2026-10-17-tsinghua-open-sources-agent-os-83-lower-task-cost/"
original_url: "https://www.qbitai.com/2026/10/503935.html"
tier: "C"
tier_name: "Weak signal"
tier_reason: "QbitAI news write-up relaying the THUNLP and ModelBest team's own release claims for PilotDeck 2.0; it adds no independent test and gives no detail on the benchmark tasks."
summary_basis: "full-text"
original_minutes: "5"
tags: ["ai technology", "model routing", "agent os", "cost"]
---

# Tsinghua open-sources agent OS, 83% lower task cost

Original title: 清华开源！智能体操作系统任务成本降83%，AI记忆可查可改

Tier C (Weak signal): QbitAI news write-up relaying the THUNLP and ModelBest team's own release claims for PilotDeck 2.0; it adds no independent test and gives no detail on the benchmark tasks.

Summary based on the full text.

## In short

A QbitAI (量子位) article (17 October 2026) on PilotDeck 2.0, an open-source "agent operating system" from Tsinghua University's THUNLP lab with ModelBest (面壁智能) and OpenBMB. It isolates projects into separate workspaces, makes agent memory visible and editable, runs long tasks in the background, and routes work between a strong planning model and cheaper executing models. The headline is a cost cut of close to five sixths on long tasks at a similar score.

## Key numbers

- 70.6 at $3.15: score and cost on seven long tasks with Sonnet 4.6 planning and MiniMax-M2.7 executing
- 69.1 at $18.36: score and cost for Claude Sonnet 4.6 alone on the same seven tasks
- $12.58 to $2.83: cost of one social-media operations run, all Opus 4.5 versus Opus 4.5 planner with Sonnet 4.5 sub-agent
- 79.1 vs 77.3: Workbuddy Bench score, PilotDeck with Qwen3.6-35B-A3B versus Claude Code

## What it reports

- WorkSpace isolation: each project gets its own file system, memory and skills, so contexts do not leak between projects.
- "White-box" memory: users can see how memories are generated, extracted, stored and used, and can edit, delete or pin them. A "Dream mode" consolidates and de-duplicates memories when idle, with one-click rollback.
- Always-on mode keeps tasks running after the chat window closes and writes results and a progress report to local files. Native MCP support; the same behaviour from web, CLI, WeCom or Feishu.
- Routing results reported by the team: on seven hard long tasks (multilingual podcast push, multi-source reports, literature reviews, codebase architecture docs and others), MiniMax-M2.7 alone scored 37.1 at $1.90; Sonnet 4.6 alone 69.1 at $18.36; Sonnet 4.6 plus MiniMax-M2.7 70.6 at $3.15.
- In a social-media operations test, running everything on Opus 4.5 cost $12.58 per run; routing polishing to Sonnet 4.5 cut it to $2.83, about 70% less.
- AGPL-3.0 licence; macOS, Linux, Windows and Docker.

## How they got it

Figures come from the development team's own tests, as relayed by the article. Scoring method, task definitions and number of runs are not described.

## Limits

- A news relay of team-reported numbers; the primary sources are the GitHub repository and an OpenBMB post, which were not reviewed here.
- Seven tasks, no variance or repeat runs reported.
- AGPL-3.0 has obligations that matter for commercial deployment.

## Why it matters for companies

The pattern is the useful part: a strong model plans, cheaper models do the routine steps, and cost drops several-fold with little loss on the team's own tests. Editable memory with rollback addresses a real governance gap, since most agent memory today cannot be inspected or corrected. Both ideas are worth testing in-house before taking the percentages at face value.

## Key passages in translation

> 从记忆的生成、抽取、存储到最后被调用，整条链路完全对用户可见。

From the generation, extraction and storage of memories to their final use, the whole chain is fully visible to the user.

> 采用主Sonnet 4.6+子MiniMax-M2.7的主强子弱架构，综合得分飙到了70.6，成本只要$3.15。

Using a strong-main, weak-sub architecture with Sonnet 4.6 as the main model and MiniMax-M2.7 as the sub-model, the overall score rose to 70.6 at a cost of only $3.15.

> 复杂的规划决策让顶配大模型来做，底层的执行、润色、格式处理全部降级给轻量模型。

Complex planning decisions go to the top large model, while low-level execution, polishing and formatting are all handed down to lightweight models.

Original: https://www.qbitai.com/2026/10/503935.html