anne hecht

The industry has moved from simple chat interfaces to agents that maintain context, take autonomous actions, and self-evolve by acquiring new skills — unlocking enterprise ROI as agents become digital employees with judgment and domain expertise.

Enterprises are moving beyond chatbots to agents that maintain context (persistent), take actions autonomously with permissions, and self-evolve by acquiring new skills (e.g., PowerPoint generation, CUDA-X libraries), driving massive token growth (4T tokens/day at Nvidia) and measurable ROI (28% of enterprises seeing immediate sustainable returns).

Security must be designed into the agent stack from the ground up: each agent runs in its own sandbox (OpenShell) with policies enforced outside the sandbox so agents cannot modify their own guardrails, while confidential computing protects model weights and data in use on GPUs.

Security for autonomous agents demands a secure-by-design stack: each agent in its own sandbox (OpenShell open-source project with SAP, Google, Red Hat contributors), policies kept outside sandboxes to prevent agent manipulation, and confidential computing extended to GPUs to protect model weights and data in-use — a capability Nvidia pioneered for accelerated computing.

Enterprise adoption is driven by CFO focus on token economics (cloud, on-prem, desktop via GB10/DGX Station) and IT governance of shadow AI; productivity gains come from sharing agent skills across teams, but require business process review to operationalize.

CFOs and infrastructure teams are optimizing marginal token costs across a hybrid deployment model — cloud, data center, colo, and edge (GB10/DGX Station for power users) — while sharing best-practice skills across the organization to improve token efficiency.

NVIDIA has pioneered confidential computing for accelerated computing; Google uses it for Gemini on-prem with Dell, Apple for consumer services. This becomes essential as enterprises run continuous agent workloads on GPUs and need to protect IP and data during inference.

The winning strategy is not frontier vs open but a blend: open models allow domain-specific post-training for accuracy and cost efficiency, while frontier models provide general capabilities; open collaboration (e.g., OpenShell, Open Secure AI Alliance) accelerates security innovation across the stack.