Eight Korean institutions will share 3.26 billion won a year to field-test medical AI agents

Divide 14.65 billion won by four and a half years and CHAIN runs at about 3.26 billion won a year ^[6][17]. Divide again across the eight participating institutions and each averages roughly 1.83 billion won for the whole award, or about 407 million won a year ^[18]. Six of the eight are university hospitals; the rest are AI companies, named in the announcement only as a category ^[7].

That is team-sized money, and the programme describes itself accordingly: building and field-testing an agentic AI orchestration ecosystem for regional, essential and public medical services ^[4]. The platform is KoreaHealth.ai, an agent system meant to plan and carry out tasks on its own ^[8]. Korean ARPA-H funds high-risk, high-reward work on health care problems conventional approaches have struggled to solve ^[3].

The output target is the more interesting term. Set 1,000 verified agents against the full award and the figure is 14.65 million won an agent ^[19], and the same money also pays for the platform and the field test ^[8][4]. Whether an agent is a narrow piece of clinical automation or a substantial one decides whether that number is generous or impossible, and the announcement does not define the unit.

The other two awards put the hospital inside other teams' programmes. One develops N-of-many gene editing therapies from domestic clinical data, with Seoul National University Hospital refining editing tools using rare-disease patient data and building a platform that links diagnosis to treatment ^[14]. In NEXUS, led by Pusan National University Hospital, it helps build a multi-institutional standard data model for cancer and a security governance framework, and joins a demonstration of AI-based prediction of cancer recurrence risk ^[15].

Two endings fit these terms. In one, the agents stay in use after the funding ends and the ministry has a per-agent cost it can re-tender against. In the other, the 1,000 count is met by logging narrow agents to a registry and the platform, not the clinical practice, is the durable output. I think the second is likelier, because the schedule fixes a count and a date without a stated adoption measure: platform finished by the end of 2027, 1,000 agents released and distributed by 2030 ^[11]. The evidence that would change that is contractual, and it sits in how the AI companies' share is structured; the announcement does not break out the split.

Paik Nam-jong, president of Seoul National University Hospital, said: "We will build essential-care AI agent infrastructure and an ultra-precise gene therapy platform to lead core technologies for the future of health care" ^[16].