When a newsroom launches five agents, what is the actual product?

The Tianyan AI application matrix platform launched at the 2026 China International Big Data Expo in Guiyang on August 28, 2026. Guizhou Daily’s launch coverage describes four shared layers: a data foundation, model capability, knowledge base, and agent orchestration. The same coverage names five application tracks for writing, interview recording, content safety, matchmaking, and education. 1

The visible agents are useful, but the more interesting product may sit underneath them. The writing engine is reported to produce a draft in five minutes, receive an initial review in 60 seconds, and output copy, a short-video script, and poster copy. The interview tool combines recording, transcription, notes, translation, Guizhou dialect recognition, 21-language interpreting, and archiving. The content-safety system is reported to use more than 2,000 risk labels. 1

A separate report describes the publisher’s archive as 76 years of news history and 300 TB of media assets. 2 The evidence establishes a publisher-owned platform with bounded applications. It raises a sharper Agentic Media question: where does the agency sit when institutional knowledge, routing, generation, and safety checks become reusable infrastructure?

What is the core: the agent, the routing layer, or the actionable knowledge base?