For most of the modern age, companies have been designed around business processes. Managers mapped how work moved from one team to another, from request to approval, from customer problem to service resolution, from opportunity to invoice. The process was the operating logic of the firm: a way of coordinating people, systems, controls and time. But the arrival of AI agents is beginning to change that picture. Jensen Huang recently put the shift starkly, arguing that companies now built around business processes may, in the future, be built on “frameworks” — proprietary systems that represent a company’s workflows and make their components more autonomous, proactive and efficient.
The word many engineers now use for this is a harness. A harness is the operating wrapper around an AI model. The model may be the source of reasoning, language and prediction, but the harness is what allows it to do useful work. It gives the model instructions, context, tools, memory, permissions, checks and boundaries. It decides which systems the agent can access, what information it can see, which actions it can take, when it must stop, and when a human needs to intervene. In plain terms, the harness is the difference between a clever chatbot and a working digital colleague. Databricks describes the harness as the infrastructure that connects a model to tools, systems, memory and execution environments, allowing it to act on tasks rather than merely respond to prompts.
This matters because enterprise AI is moving from conversation to action. The first wave of generative AI helped people draft emails, summarize documents and search knowledge bases. Useful, but still largely assistive. The next wave is different. Agents can inspect records, call software tools, write code, check their own work, escalate exceptions and complete multi-step tasks. Once AI begins acting inside the operating fabric of the company, the key question is no longer simply whether the model is smart. It is whether the surrounding system is reliable, observable, governed and aligned with the way the business actually creates value. That is why harness engineering is becoming important: it turns general intelligence into situated capability.
NVIDIA’s view of this shift is strategic. Huang’s claim is not just that agents will automate old workflows. It is that every company has its own domain-specific wisdom, and that this wisdom should be embodied in proprietary agentic systems rather than outsourced entirely to generic AI providers. In the same discussion, he argued that the most valuable enterprise agents will combine open models, company knowledge, tools, memory, guardrails and runtime environments. He also described these specialized agents as among a company’s most precious assets.
LangChain fits naturally into this picture. It provides software infrastructure for building agents around models: connecting them to tools, managing context, supporting memory, orchestrating long-running work and making agent behavior easier to evaluate. Its recent collaboration with NVIDIA around the NemoClaw Deep Agents Blueprint is a useful example of where the market is heading. LangChain says the blueprint combines an open model layer, a tuned agent harness and a governed runtime, so enterprises can build agent systems they can tune, secure, measure and improve. It also argues that production performance depends not only on model choice, but on the entire system around the model: tools, context, evaluation, runtime and policy.
In earlier phases of AI adoption, many leaders assumed the main decision was which model to use. AI models, however, are only one part of the system. LangChain and NVIDIA have shown that performance can improve not by changing the model itself, but by tuning the harness around it: how the agent uses tools, manages context and evaluates intermediate steps. The implication is simple but powerful: two companies may use similar models and get very different results because one has designed a better environment around the intelligence. In the process era, companies competed on operational excellence. In the harness era, they will compete on how effectively they turn AI capability into governed action.
You can already see this pattern in software engineering. OpenAI uses the concept of harness engineering to describe the work of shaping the environment in which coding agents operate. In practice, that means much more than writing a better prompt. It includes repository structure, documentation, architectural constraints, tests, validation loops and other machine-readable guidance that allows an agent to work productively without causing chaos. OpenAI’s approach encodes scaffolding, feedback loops, documentation and architectural constraints into artifacts that Codex agents can use across coding, testing and observability workflows.
Customer service offers another example. Salesforce now describes an AI agent harness as the operational layer that manages an agent’s tools, memory, safety, context, permissions, and interactions with business systems. In customer service, that matters because the agent is not simply answering questions; it is operating inside a relationship. Salesforce’s own discussion of Agentforce emphasizes that a good AI agent should know when to step aside, preserving context, maintaining continuity, routing the customer to the right human, and avoiding the familiar frustration of making people repeat themselves. The handoff, in other words, is not a minor feature. It is part of the design. The harness is not just automating work; it is shaping the relationship between machine speed and human care.
Microsoft’s strategy points to the same shift, but with a sharper concern: who owns the intelligence that agents create and absorb? Satya Nadella has argued that companies should not let their harness, context and tacit knowledge become captive to a single frontier model; if a firm’s expertise lives in its people and workflows, some of that intelligence also needs to live in systems the firm controls. Microsoft’s response is to help companies build the system around the model: a governed environment where enterprise knowledge, tools, workflows, permissions, feedback and human oversight can accumulate into a proprietary harness. Its agent platform strategy is built around this idea — agents should be contextualized in enterprise data, governed by design, observed in production, improved through feedback, and run under the company’s control. In this view, the old workflow is no longer just a diagram showing how work should move. It becomes an owned intelligence system — one that must be built, improved and defended.
This is why the comparison with business processes is so important. A business process describes how work should move through an organization. A harness defines how an intelligent system should act inside that organization. The process routes tasks; the harness gives agents context, tools, memory, permissions and feedback. The process assumes a relatively stable sequence of steps; the harness allows adaptive, multi-step execution. The process is usually managed through ownership, compliance and continuous improvement. The harness must also be managed through evaluation, observability, policy, access control and human oversight. A process is a map of work. A harness is an operating environment for intelligent work.
But there is a danger in stopping there. The language of harnesses is powerful, but it remains incomplete. It can tempt leaders to take existing workflows, wrap them in agents and call that transformation. Yet a faster workflow is not necessarily a better one. A poorly designed process can become more brittle when automated. A weak approval system can become more dangerous when accelerated. A bad customer experience can become more alienating when scaled. A biased pattern can become more entrenched when embedded in software. The harness answers an engineering question: how do we make this agent work? Business leaders must ask a larger question: how do we design the systems in which great decisions happen?
This is what we should think of as cognitive design. Every organization now has to decide how intelligence should be distributed between people and machines. Some activities benefit from machine speed, scale and pattern recognition. Others depend on human judgment, empathy, accountability, creativity or moral reasoning. Most of the important work will sit somewhere between the two.
The leadership challenge is not to automate as much as possible, but to place human and machine intelligence where each creates the greatest value and the least hidden risk. In this sense, AI transformation is not simply a technology rollout. It is a rethink of how the enterprise thinks, senses, decides and acts. That’s no small task by the way. This is the question my co-author and I explore in Abundant Intelligence: How Digital Labor Will Rewrite the Rules of Business, where we set out a methodology for designing the relationship between human and machine intelligence.
In our view, the winners of the next phase will not necessarily be the companies with the most agents or the most compute. They will be the companies with the best cognitive architecture: the clearest view of where humans should lead, where machines should act, where oversight should sit, and how trust should be preserved as work becomes more autonomous. Harnesses will be essential to that architecture, but they are only one layer. They make agents useful. They do not, by themselves, determine whether the company has placed intelligence wisely.
Huang is right that companies are moving beyond business processes toward frameworks and harnesses. But the final twist is that the most important harness may not be the software wrapper around the model. It may be the organization itself. The company supplies the purpose, context, constraints, memory, values and accountability within which intelligence operates. Business processes told people how work should move. Harnesses tell agents how work should be done. Cognitive design asks the deeper question: how should the enterprise think?
The future company will not just have an operating system. It will have a mind.