Chinese retailer and technology company Alibaba Group Holding Limited (Alibaba) is set to announce a new agentic AI tool built on its Qwen large language model as soon as this week, Bloomberg has reported.

The product is designed to compete with popular mass-market AI agents like OpenClaw – which can be connected to services such as WhatsApp and Slack – while being specialised for enterprises, insiders said.

Alibaba plans to integrate the agent with other services it provides over time, such as its Chinese retailer Taobao and its FinTech platform Alipay. It is being developed by the team behind the company’s workspace platform DingTalk, according to people familiar with the matter.

Those people also said the tool can help firms operate computers, browsers and cloud servers, and has built-in features to safeguard data security.

Alibaba’s Qwen 3.5, launched in mid-February, claims to operate more cheaply and at higher throughput than its US competitors including GPT-5.2 and Gemini 3 Pro. A campaign run by the company in the same month drove a sevenfold increase in active users.

The company is expanding its reach in the AI sector across a number of channels, having committed to invest at least RMB 380 billion (US$53 billion) over the next three years.

Alibaba’s agentic AI push has not been without problems, however. On 6 March, machine learning researcher Alexander Long noticed a short passage in a paper released on 31 December 2025 that revealed Alibaba’s experimental AI agent ROME had displayed unintended behaviours during training. His post on X covering the paper has since amassed over 2.8 million views.

Most strikingly, the agent quietly diverted GPU power from Alibaba toward cryptocurrency mining, despite never being instructed to do so. It did so by establishing a “reverse SSH tunnel from an Alibaba cloud instance to an external IP address,” the researchers said. The decision was noticed by Alibaba’s internal firewall rather than the research team working on ROME.

---