Nvidia has released a design for running AI agents inside fixed limits, and it is available now.
The company announced the OpenShell Reference Architecture at its GTC Berlin conference. It covers setups from enterprise infrastructure to edge machines, meaning computers that sit outside a data center and close to where the work happens. Nvidia said Anthropic, OpenAI, Meta, Microsoft, Google Cloud and AWS are among the companies working with it. That partner list is Nvidia's own description of the effort.
Agents are AI software that takes actions, such as running code or changing files, instead of only answering questions. Once software acts, the useful question shifts from what it says to what it is allowed to do. A set of fixed limits is one answer to that.
A reference architecture is a published blueprint that other companies build from. It is not a product you buy, so its value depends on how many people adopt it. Anthropic and OpenAI compete directly, so any shared list that includes both is notable.
The open question is whether names on a partner list turn into the same limits being applied across rival systems, or whether each company keeps its own version of the fence.