Crowdworks said on Wednesday it won an "AI agent safety and reliability verification framework support" project ordered by the Ministry of Science and ICT and the National Information Society Agency (NIA).

The project will be promoted to build a verification framework that can assess the safety and reliability of AI agents as they draw up plans and call external tools to carry out tasks.

Crowdworks will be responsible within the consortium for building multi-step reasoning scenarios and a Korea-specific verification dataset. The goal is to enable evaluation not only of the accuracy of final answers but also of the validity of the process by which the agent derives answers and performs tasks.

To that end, it plans to design scenarios by difficulty that can be solved only by sequential or parallel calls to multiple Model Context Protocol (MCP) tools and external APIs. It also plans to build more than 7,000 verification dataset entries reflecting domestic service environments and changing data.

SureSoftTech will lead the project, with participation from Crowdworks, the Korea Advanced Institute of Science and Technology (KAIST) and the Korea AI Industry Association. The consortium plans to complete construction of the AI agent verification framework by the end of this year.

A Crowdworks official said, "Safety and reliability verification of the execution process of AI agents is becoming important." The official added, "We will contribute to establishing a verification framework suited to the domestic environment by leveraging our data-building and evaluation capabilities."