Whether software that runs office work on your behalf can move from trial into production turns on something other than how clever it is.
Cohere and OpenText announced a partnership. The goal is to move agents — software that takes an instruction, works out its own steps, and operates one internal system after another — from pilots into production inside governments and heavily regulated industries. Cohere's enterprise models sit on OpenText's corporate data layer so the software can read internal records and work across several systems. Deployment can be chosen: a company's own hardware, or a cloud kept within national borders. That choice is the condition for handling data that cannot leave the building. NVIDIA, separately, released a way to hand an agent the job of rewriting GPU routines in Rust. The translation runs as a pipeline with machine checks at every stage, and NVIDIA says it moved all 24 published operators across.
The two answers are not the same one. What Cohere moved is the boundary — what the software is shown, and where it is allowed to sit. What NVIDIA moved is the stage after: the output is filtered by machine before a person reads it, and only clears once its speed has been measured too. Decide the scope up front, or check everything that comes out the back. Putting this software where mistakes are not tolerated, what each built first was not intelligence but one of those two enclosures. That said, the Cohere announcement is a framework: which jobs actually went live is not stated.
Which task the Cohere–OpenText pairing puts into production first. For NVIDIA, whether the same checks hold outside the published operator set — on the idiosyncratic code sitting inside real codebases. That is where an enclosure either proves out or stays on paper.