Is an AI agent just an LLM plus a harness?

Nvidia is promoting a modular vision for AI agents, separating the model from its "harness." The definition has strategic implications for where infrastructure and product value will accrue.

Where it happened

The framework was detailed in a June 2026 interview with Nader Khalil, Nvidia's Director of Developer Technologies, published in The New Stack. The interview outlined Nvidia's internal thinking and its investment in open-source agent tooling like OpenClaw.

Side A: The agent is a modular system

Nvidia's position, articulated by Khalil, is that an agent is best understood as two distinct parts. "An agent is an LLM and a harness," he stated. In this view, the Large Language Model is a powerful but raw component, like a CPU. The "harness" is everything else: the control loop, memory systems, tool integrations, security policies, and context management.

Proponents of this view argue that enterprise-grade reliability and product differentiation happen in the harness, not the model. The evolution from simple system prompts to complex, multi-tool integrations is seen as a story of improving the harness. Nvidia is acting on this thesis by contributing full-time developers to the OpenClaw agent framework and providing NemoClaw, an enterprise reference architecture for deploying these systems on Nvidia hardware. The strategy treats the LLM as a powerful, swappable reasoning engine, while the stable, defensible value is built in the surrounding infrastructure.

Side B: The agent is the model

The alternative view holds that the harness is a temporary scaffold, not a permanent architectural layer. This model-centric perspective argues that the distinction between "LLM" and "harness" is an artifact of current model limitations. Future foundation models, proponents believe, will natively integrate capabilities like long-term planning, memory, and tool use, making complex external harnesses obsolete.

According to this thesis, investing heavily in elaborate harness infrastructure is a bet against fundamental progress in AI research. The ultimate goal is not a more sophisticated harness but a model that no longer requires one. Every function absorbed into the model from the harness represents a step forward. From this perspective, the harness is a collection of features that are simply waiting to be subsumed by the next generation of more capable, inherently agentic models.

What's underneath

This is not just a technical debate about architecture; it is a strategic disagreement about where value will be captured in the agent economy. The modular "harness" view, championed by an infrastructure provider like Nvidia, creates a stable middleware layer where standards can be set and ecosystems can be built, all driving demand for the underlying compute. It carves out a durable, defensible space for tooling and platform companies. The model-centric view, conversely, suggests value will continue to concentrate in the foundation models themselves, with the companies training the largest models absorbing the agent-level capabilities over time. The two sides represent different bets on the pace and direction of AI progress.

The investor read

Nvidia's 'LLM + harness' thesis is a direct attempt to shape the AI agent infrastructure market in its favor, akin to its CUDA strategy for GPUs. By defining the 'harness' as a distinct, critical layer, Nvidia encourages the growth of a tooling ecosystem that runs best on its hardware. This signals a potential commoditization of core LLM reasoning, with significant value accruing to the companies that build the standardized harnesses, control planes, and security wrappers for enterprise agents. For investors, this frames the agent stack as a key battleground, with opportunities in the emerging 'harness' layer, not just in the foundation models themselves.

Pull quote: “An agent is an LLM and a harness.”