Alibaba Cloud's community blog published a write-up on 28 October of Corpus Formation Autopilot, the Track 4 autopilot-agent winner in its Global AI Hackathon Series with Qwen Cloud.
The project takes a plain-language description of a business, has Qwen call tools that search a US law library, classify the business activity and prepare a formation handoff, and returns the results for the user to review. Qwen-flash and qwen3.7-max run on Alibaba Cloud Function Compute. The Apache-2.0 Corpus Agent Kit connects the model to those tools and ships two more entry points: an MCP bridge so compatible assistants can reach the same toolkit, and an embeddable widget.
The design keeps irreversible actions out of the demo. The demo cannot charge a card or file a company; payment and filing live in a separate hosted workflow that the user enters deliberately. The authors also state the limits plainly: legal coverage varies by jurisdiction, and generated guidance should be checked against the underlying sources.
Our takeSeparating "the model reasons" from "the model acts" is the part worth taking from a hackathon entry. Tools return grounded text, the model picks which lookup to run, and anything that spends money or files a document requires a person to walk through a different door. That is a stronger guarantee than a prompt telling a model not to do something. The residual risk sits elsewhere: a classifier that mislabels the business activity sends the user down the wrong legal branch, and nothing here catches that.