Alibaba rebuilds Qoder as coding agent for non-developers

On Alibaba’s rebuilt Qoder, a user can describe a software idea in conversation and wait while the code work happens out of sight. Unveiled on August 27, the coding agent is being repositioned as a “universal workbench” for non-developers, not just programmers: product, design, operations, data and independent creator roles are all in its target audience.

The mechanism is deliberately conversational. Qoder breaks a request into requirements, refines them through chat, generates and executes code, then checks the result, according to Alibaba. Its year-old Agent Harness foundation is designed to handle long-running tasks and decompose complex work. Users can also issue instructions through real-time voice control while the assistant tracks tasks in progress.

There are two modes. Developers can select models and control the pipeline themselves. In general mode, an auto-scheduling engine makes the trade-off between output quality, speed and cost without requiring manual tuning. Qoder also connects to more than 40 connectors, 70 plugins and a 20,000-skill library spanning code repositories, project-management tools, cloud services and internal business systems.

Alibaba says Qoder reached 100,000 users in five days, passed one million in six weeks and generated $100 million in revenue within eight months of its first launch. It now reports more than six million total users and over 100,000 enterprise customers. The figures come from the company, which presents them as evidence that coding agents are moving toward broader work automation rather than code generation alone.

So what changes in practice? A team member without programming skills could move from a spoken or written idea to a usable software result through an iterative chat, while an experienced developer gets more control over the same underlying system. Alibaba is offering 500 Qwen model credits to individual users on five consecutive days through September 1 and a 50 percent discount for enterprises using Qwen3.8-series models, but the materials give no broader pricing or independent assessment of the outputs.