Distyl, NVIDIA, and the Reality of Enterprise Agents
What production AI actually demands-and the machinery we are building to deliver it.
We are past the honeymoon phase of the generative AI revolution. The pilot programs are finished. What remains is the rigorous work of making it real. Today, Distyl is announcing integration of NVIDIA AI Enterprise software into Distillery, our enterprise AI platform.
Access to a smarter model is merely the baseline requirement. The fundamental challenge of this next phase is operationalizing that intelligence at scale, and doing so without compromising the stability of the organization. A true autonomous enterprise agent must reason over complex states, execute generated code, spawn sub-agents, and operate deep within sensitive internal systems. It must do all of this while navigating the strict, unforgiving realities of corporate governance and compliance. In practice, that requires systems built for long-running enterprise assistants, agents that persist across interactions, accumulate context over time, and continuously reason over evolving enterprise data rather than responding to isolated prompts.
That exacting reality shapes how we build at Distyl.
When you examine the architecture of most enterprise AI efforts, you realize they are missing two crucial ingredients.
The first is context. You can deploy the most sophisticated model on earth, but if you drop it into a corporate intranet, it operates blind. The organizational knowledge it needs is present, yet trapped in silos. It sits behind access controls never designed for a machine, and lives in formats that defy programmatic traversal. Without a structured way to ingest this environment, agents hallucinate. They operate from a fractured, incomplete view of the world.
The second missing piece is enterprise-grade controls. In the real world, you need evaluations, versioning, monitoring, governance, and coordination across many agents handling both happy paths and failure cases. No matter how intelligent an agent is, without the proper enterprise-grade harness, it will never get into production.
Distillery is our answer to these problems. At its core is what we call the Context Mesh. We moved past the concept of the flat, static retrieval index and built a living, traversable graph of an organization’s nervous system. This structure encompasses episodic memory, tacit and explicit knowledge, policies, workflows, and domain logic. Agents navigate this mesh deliberately. They survey available data, activate specific business routines, and write back state that persists long after the session ends. Backing all of this is a complete set of enterprise-grade controls.
This persistent context layer is what enables long-running enterprise assistants. Instead of resetting on every interaction, agents accumulate memory, understand entities and relationships over time, and maintain continuity across workflows, decisions, and customer interactions.
At Distyl, we believe that different runtime choices, and models, are right for different enterprises, and such support different runtime options for agentic applications. To harness the powers of NVIDIA’s latest technology capabilities, we built a personalized assistant that leverages NVIDIA’s enterprise offerings. For runtime, we used the NVIDIA NeMo Agent Toolkit. It provides the declarative configuration and plugin architecture needed to orchestrate entirely different operational patterns on a single, unified foundation. For interactive, unpredictable work, we deploy a ReAct agent with full tool-calling autonomy. For tasks that demand meticulous, grounded investigation, we switch tracks to a deterministic deep-research pipeline based on NVIDIA AI-Q Deep Agent Blueprint. Here, the code lays the tracks and drives the execution path, while the model does the heavy lifting of reasoning over the results at every single stage.
To support long-running agents in production, we’re also working together on NVIDIA NemoClaw, an open-source stack that simplifies running OpenClaw long-running assistants safely in production. As part of the NVIDIA Agent Toolkit, NemoClaw installs the NVIDIA OpenShell runtime, a secure execution environment for autonomous agents, along with open source models such as NVIDIA Nemotron. This architecture makes it significantly easier for enterprises to deploy always-on assistants that can safely execute code, orchestrate tools, and operate continuously across enterprise systems.
The engine driving this entire apparatus is open NVIDIA Nemotron 3 Super, deployed via NVIDIA NIM microservices. It features 120 billion parameters with 12 billion active at inference, a massive 1-million-token context window, and five times the throughput of its predecessor. This is critical for complex agents that coordinate across tens of tools, and thousands of pieces of context to carry out enterprise operational workflows with high precision.
We are very excited about how NVIDIA Nemotron 3 Super integrates with the broader NVIDIA AI Enterprise software stack. We see strong customer interest in open source models for cost, latency, and governance reasons, and it’s exciting to offer these capabilities on Distillery.