From OpenClaw to NemoClaw: The Evolution of AI Agent Accountability

By Carole Achramowicz, Vice President, Product Marketing

We’re in the midst of the biggest technological shift since ChatGPT, and it has already begun revolutionizing everything from the way that we manage our daily calendars and track the news, to the fundamental ways that work gets done. In his recent GTC keynote, NVIDIA CEO Jensen Huang described this technological leap as moving from a world of “Assisted AI,” where models suggest content or code, to a world of “Agentic AI,” where autonomous agents independently execute complex tasks across enterprise systems.

And this shift is happening with unprecedented speed. In our recent Jitterbit AI Automation Benchmark Report, we found that only 1.6% of respondents have no AI agents deployed. For those with agentic deployments, the number of agents is set to increase an average of 43% over the coming year — with the percentage of respondents with massive, 100+ agent deployments set to more than double.

This meteoric growth in these agents has led to concerns about agent sprawl and agentic management, spurring industry excitement around frameworks like OpenClaw (an open-source system capable of connecting messaging apps like WhatsApp or Discord to autonomous agents running on a user’s machine.)

But while the shift from agent capability to agent management is an important one, as organizations transition from AI pilots to full-scale operations, a more forward-looking approach is required.

The Tipping Point: From Inference to Agency

Traditional AI models focus primarily on inference, or interpreting data and supporting human decisions. Agentic AI, on the other hand, fundamentally changes the role of software: Instead of humans directly manipulating tools, applications are becoming environments where agents operate on the user’s behalf.

This represents a tremendous leap in terms of capabilities: It unlocks the ability to turn large language models into autonomous actors capable of navigating applications and coordinating workflows across disparate data sources. But it also magnifies risk, opening the door to compliance violations, agents getting stuck in loops, permissions creep and more. Responsible, future-proof deployments require us to expand our approach to include the foundational requirements of trust, control and accountability.

Why Security Layers and Governance Matter

In introducing NemoClaw, NVIDIA’s software stack and plugin that adds enterprise-grade sandboxing and policy-driven controls to OpenClaw, Huang described the new offering as a way to allow LLMs to interact with software through “computer use” capabilities. This opens the door for AI to operate browsers and applications just like humans do.

This transition represents a critical crossroads for the modern enterprise, according to Jitterbit CEO Bill Conner:

“We’re seeing one of the most rapid and monumental technology inflection points since ChatGPT: the shift from AI that suggests to AI that acts. OpenClaw accelerated that transition, but with that power comes real risk. In its raw form, these frameworks lack an inherent security model. In an enterprise environment, that’s not just a gap, it’s a liability.”

As AI agents gain the ability to move dynamically across systems and jurisdictional boundaries, the challenges of data sovereignty and accountability become increasingly urgent. That’s why Jitterbit advocates for a “layered” approach to AI. By wrapping agent frameworks with a layer of robust infrastructure, guardrails and policy enforcement, enterprises can capitalize on the efficiency of AI without compromising control.

According to Jitterbit CTO Manoj Chaudhary, NVIDIA’s newest offering suggests a sea change in the way AI agents are allowed to operate. “The introduction of NemoClaw reflects a broader industry realization: Agentic AI needs an enterprise-grade foundation. We are moving toward ‘AI operating systems’ — platforms that don’t just enable intelligence, but also control how that intelligence behaves. This is open innovation at the edge, with structured governance layered on top.”

Moving Toward Agent-as-a-Service

The secure democratization of data access, Conner said, requires two foundational elements: human oversight and technological guardrails.

“As agents become more autonomous, AI accountability becomes harder to define. When an agent acts independently, responsibility must still be clearly assigned to the system, the organization, and ultimately to human oversight. Guardrails must evolve from simple policy controls into dynamic, enforceable constraints embedded directly into how agents operate.”

In other words, while the “deploy now and ask questions later” approach might have gotten us this far, it’s becoming increasingly clear that it won’t take us where we actually need to be. In the race to agentic adoption, the enterprises that ultimately cross the finish line won’t be the ones with the most powerful agents, but the most accountable ones.

As AI agents continue to evolve into the backbone of enterprise operations — toward a future Huang describes as “AI for everything” — the conversation must continue to shift from AI agent capabilities to AI agent accountability. The ultimate goal is to ensure that agents are not just powerful, but also safe, compliant and integrated into the fabric of business operations.