Event date · · Alibaba

Qwen releases Qwen-Live-Delivery-Plugin for human-in-the-loop long-running AI agents

Alibaba 阿里巴巴Chinese AIOpen weights
FACT STATEMENT

Qwen (Alibaba) released Qwen-Live-Delivery-Plugin 0.1.0, an Apache-2.0 licensed TypeScript plugin for Codex CLI that lets long-running AI agents deliver iterative versions while users steer work and adjust budgets. It supports Linux, macOS, and Node.js 22+, and was tested with Codex 0.166.0/0.167.1.

China context

Original name
Qwen-Live-Delivery-Plugin
Outside China
Open weights · github.com
Claims
Company-reported; not yet independently evaluated
For builders
Developers outside China can install the plugin from GitHub and use it with Codex CLI; it is open source under Apache-2.0 and supports Linux, macOS, and Node.js 22+.
For investors
This release shows Qwen's continued investment in agent tooling and developer experience, potentially strengthening its position in the open-source AI ecosystem.
What happened

Qwen-Live-Delivery-Plugin is a lightweight plugin for long-running AI agent tasks. It includes a Live Delivery Skill, an MCP server, and lifecycle hooks that enable human-in-the-loop control: users can add requirements, adjust time budgets, preview and download versions, and stop tasks while delivered versions remain available. The plugin works with Codex CLI and provides a local browser workbench for monitoring and interaction.

Technical significance

The plugin uses three modules: a Live Delivery Skill ($ld:go) that guides the agent to produce a minimal usable first version and iterate based on results; a Live Delivery MCP with commands ld begin, ld publish, ld status, and ld done to manage delivery plans and state; and lifecycle hooks that inject user feedback into context and block premature turn completion. It supports browser-created tasks and external CLI tasks, with sandboxing defaults to workspace-write unless explicitly overridden.

Industry impact

Developers using Codex CLI gain a way to supervise long-running agent tasks without losing delivered work, reducing the risk of wasted compute and unusable outputs. This addresses a control gap for agentic workflows and may pressure other agent tooling providers to offer similar human-in-the-loop delivery mechanisms.

Decision value

For teams building on Codex, this plugin can lower the cost of long-horizon agent tasks by enabling incremental delivery and user steering, potentially reducing rework and improving output quality. It is open source under Apache-2.0, so there is no licensing fee, but it requires Codex and its associated model/API costs.

What to watch

The repository lists planned support for additional agent harnesses beyond Codex, with end-to-end compatibility tests for release publishing, mid-task requirement changes, session continuation, and cancellation. Watch for commits adding harness integrations and test coverage as the next signal of broader adoption.

Latest in Chinese AI

  1. HuaweiReportedly: Huawei Cloud launches open-source Pangu model openPangu-2.1-Pro on MaaS platform
  2. Infinigence AIReportedly: D-Robotics and Infinigence AI Bring RLinf and Three Other Open-Source Robot Tools to Sunrise Chips
  3. Moore ThreadsReportedly: Tsinghua University releases ProtCell single-cell proteomics foundation model, reproduces pretraining on Moore Threads MTT S5000
  4. TencentReportedly: Tencent Hunyuan releases real-time speech model Hy Realtime with tool calling and singing, live in Yuanbao voice calls and Tencent customer service
  5. MeituanMeituan LongCat open-sources LongCat-Video-Avatar-1.6 for audio-driven human video generation

All China AI Events

AIGC Newsletter

China AI, with sources and context.

Analysis of Chinese AI models, companies and policy, and what you can use outside China.