SIAgentNews.com
← All news

Alibaba open-sources Qwen Code 1.0 preview with new agent harness

Alibaba merged its Qoder coding product's agent framework into the open-source Qwen Code project to help developers build more scalable AI coding tools.

· First reported by pandaily.com

1.0.0-previewCurrent preview version

The short version

  • Qwen Code 1.0 Preview merges the 'harness' from Alibaba's Qoder product into an open-source repository.
  • The new architecture uses an in-memory event bus to separate reasoning orchestration from tool execution.
  • The update allows for distributed operations across multiple nodes rather than hosting sessions in a single process.
  • Developers can now embed the harness into their own products under the Apache-2.0 license.
  • The preview version is hosted on a separate GitHub branch to keep the existing stable version intact.

Why it matters

This move makes it easier for developers to build complex AI agents that can actually perform tasks, not just generate text. By open-sourcing the 'harness,' Alibaba is providing the underlying plumbing needed to create scalable, production-ready coding assistants.

Merging the agent harness

Alibaba recently announced the release of Qwen Code 1.0 Preview, a significant update to its open-source coding agent. The company chose to fold the 'harness' from its Qoder coding product into this new project. The harness acts as the essential layer that connects a model to its available tools and the actual execution process.

By making this component open-source, Alibaba is allowing other developers to embed the same framework into their own software products. This means that instead of building a tool-use system from scratch, developers can use Alibaba's existing architecture. The project is currently available under the Apache-2.0 license, which is a common open-source license.

The new preview version is hosted on a separate branch of the QwenLM/qwen-code repository. This structure allows the community to test the new features while the existing stable branch remains unaffected. As pandaily.com reported, the project now lists a v1.0.0-preview pre-release alongside the current 0.27 line.

A new distributed architecture

The core of the 1.0 update is a shift toward a more scalable, distributed operation model. It starts with an in-process, in-memory event bus that extends a unified event model. This design is intended to move away from a system where one process hosts several sessions. Instead, it allows multiple nodes to host many more sessions simultaneously.

A key technical concept in this update is the 'brain and hand separation.' This means that the reasoning orchestration, which is the 'brain' of the agent, is kept under independent management from the tool execution, which acts as the 'hand.' By separating these two functions, the system can manage complex tasks more effectively without the parts interfering with one another.

The architecture organizes intake, scheduling, and execution into distinct layers. This modularity ensures that the main execution flow remains relatively undisturbed. It allows the system to handle high-volume requests by distributing the workload across different hardware resources more efficiently than previous versions.

Practical effects for developers

Alibaba identifies four primary benefits for those using the new Qwen Code 1.0 Preview. First, the processing units for reasoning, memory, and tools can be plugged in around events with minimal intrusion. This modularity makes it easier for developers to swap out different components depending on their specific needs for a coding assistant.

Second, the model access is centralized through a single entry point. This central point manages configuration, authorization, and call records in one place. Third, the system manages the Model Context Protocol (MCP) for connecting tools in a unified way. This simplifies how the agent discovers and interacts with various external software tools and APIs.

Finally, the update improves how sandbox resources are allocated and reused. Resources are now managed within the same configuration and permission boundaries. Alibaba states this should reduce the need for repeated initialization and minimize idle usage, which ultimately helps to lower the total running costs for developers hosting the model.

Integrations and availability

Qwen Code is designed as an open-source coding agent that primarily runs in a terminal environment. However, the repository shows that it supports a wide range of integrations for different workflows. These include desktop, browser, and editor integrations for popular tools like VS Code, Zed, and JetBrains.

While the preview is now available, Alibaba has not yet provided a specific release date for the stable 1.0 version. They also have not published any official benchmark results to measure the performance of this new architecture. The 'preview' label serves as a notice that the interfaces and features may still undergo changes during the testing phase.

It is currently unclear how closely the open-source kernel matches the version of the harness used inside the commercial Qoder product. The performance and cost-saving claims provided by Alibaba are based on their own internal descriptions. Independent tests have not yet been conducted to verify these specific improvements in a live environment.

What to watch

  • When the official stable 1.0 version will be released.
  • Independent performance benchmarks comparing the new distributed model to previous versions.
  • How many third-party developers adopt the open-source harness for their own AI agents.

Try it

Check the QwenLM/qwen-code repository on GitHub to see the difference between the stable branch and the v1.0.0-preview branch.

Sources and links

First reported by pandaily.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news