The current enterprise AI landscape is defined by a widening gap between the theoretical capabilities of autonomous agents and the reality of their production deployment. For months, the industry narrative has focused on the race for larger context windows and higher reasoning benchmarks, assuming that a smarter model would naturally solve the problem of agency. However, as companies move from controlled pilots to full-scale operations, they are discovering that the bottleneck is not the intelligence of the model, but the rigidity of the systems the models are expected to control.

The Friction of Legacy Systems

At the VB Transform 2026 conference, infrastructure leaders from LinkedIn, Walmart, and Zendesk converged on a singular realization: AI agents are thinking in milliseconds, but the infrastructure they rely on is still operating in seconds. This temporal mismatch creates a systemic failure that no amount of prompt engineering can fix.

LinkedIn encountered this friction first within its Kubernetes environment. In a traditional container orchestration setup, the system spins up a container upon request. While this is efficient for human-triggered applications, the several-second delay required for container initialization is an eternity for an AI agent executing a multi-step chain of thought. To bridge this gap, LinkedIn abandoned on-demand provisioning in favor of pre-provisioned pools. By maintaining a ready-to-use set of containers and swapping workloads in real-time, they reduced the latency to a level that matches the agent's operational speed.

Beyond latency, LinkedIn addressed the inherent instability of LLM-to-LLM evaluation. When one model evaluates the output of another, the risk of compounding hallucinations increases. Their solution was to strip the LLM of unnecessary responsibility. LinkedIn restructured its agent workflows so that approximately 80% of the process is handled by deterministic script code. The LLM is only invoked at specific points where genuine reasoning or synthesis is required, ensuring that the core logic of the agent remains predictable and verifiable.

Walmart faced a different kind of infrastructure crisis: the chaos of the citizen developer. As internal tools made it easier for non-engineers to build their own agents, the company saw a surge in innovation. However, this democratization led to a proliferation of redundant agents, with dozens of different tools attempting to solve the same problem without coordination. Rather than restricting access to these tools, Walmart implemented a governance layer. This system detects redundancy and identifies the most effective version of an agent before promoting it to the production environment, effectively turning a management bottleneck into a curated pipeline.

Zendesk's struggle was rooted in the data layer. The company possesses a massive repository of 20 billion customer conversations. The initial instinct was to leverage the expanding context windows of frontier models, attempting to feed vast amounts of data directly into the prompt. This approach failed to produce the precision required for high-stakes customer service. Zendesk realized that the context window is not a substitute for a robust data architecture. They shifted their investment away from model-dependency and toward the underlying data pipelines and infrastructure, ensuring that the agent receives the right data at the right time rather than simply receiving more data.

The Pivot to Infrastructure Independence

These three cases signal a fundamental shift in the AI adoption curve. The industry is moving away from a model-centric approach—where the choice of LLM is the primary strategic decision—toward an infrastructure-independent strategy. The goal is no longer to find the best model, but to build a system that is vendor-agnostic, allowing the model to be swapped as easily as a software library.

LinkedIn has operationalized this through the creation of an AI Gateway. This single interface acts as the entry point for all model calls, decoupling the application logic from the model provider. Complementing this is a dedicated memory subsystem that maintains context independently of the LLM. This architecture allows LinkedIn to maintain consistent API semantics across public clouds and on-premises data centers, ensuring that they can switch providers without rewriting their entire agentic framework.

Walmart has adopted a similar tiered approach to workload management via its internal gateway. They categorize tasks into three distinct streams: fully deterministic workflows for high-security and regulated tasks, planner-reasoner workflows for open-ended problem solving, and a hybrid of the two. By fixing the logic of sensitive tasks at the design stage, they remove the risk of stochastic failure. For Walmart, the choice between a frontier model and an open-weight model is no longer a matter of brand loyalty or a fixed policy, but a tactical decision based on the specific requirements of the workload.

This trend suggests that the role of frontier models is shrinking. While these models are essential for the narrow, high-complexity reasoning tasks that define the cutting edge, they are increasingly viewed as specialized components rather than the center of the stack. Zendesk's analysis confirms this, noting that enterprises increasingly want to own their models and infrastructure. The proportion of AI use cases that actually require the extreme reasoning capabilities of a frontier model is smaller than previously assumed, making the ability to run smaller, specialized models on owned infrastructure a competitive advantage.

For organizations currently deploying agents, the path to production requires a shift in priority. The focus must move from benchmarking model performance to building a robust agent harness. This means establishing a rigorous evaluation system (Evals) before the first agent is even deployed. As Zendesk emphasized, the ability to break down a problem into measurable metrics is the only way to accelerate deployment speed.

Furthermore, the ownership of the context is critical. Storing enterprise knowledge within a model provider's proprietary ecosystem creates a dangerous dependency. By maintaining context and data within their own infrastructure, companies ensure that their accumulated experience and data assets remain portable. When tomorrow's open-source model outperforms today's frontier model, the companies that own their harness and their data will be the only ones capable of pivoting instantly.