At VB Transform 2026, technology leaders from LinkedIn, Walmart, and Zendesk shared a surprising finding: the real bottlenecks for AI agents are not the models but the legacy infrastructure built for human workflows. Animesh Singh, senior director of AI platform and infrastructure at LinkedIn, Desiree Gosby, SVP of corporate technology services and technology strategy at Walmart, and Sami Ghoche, VP of applied AI at Zendesk, each described the walls they hit when moving agents from pilot to production. Every company arrived at the same conclusion from a different starting point: infrastructure designed for how people work cannot keep up when agents operate in milliseconds.
LinkedIn: Kubernetes and structural hallucinations
LinkedIn's first bottleneck was Kubernetes, not a model. Singh explained that on-demand container provisioning takes seconds, which is too slow for agents. The fix was pre-provisioned container pools that swap agentic workloads in real time. A tougher problem emerged when agents controlled their own orchestration. A five-point evaluation system seemed clean, but hallucinations persisted. Singh traced the issue to a structural flaw: an LLM evaluating another LLM's output shares the same failure mode. LinkedIn built its own harness and control flow, pushing LLMs only where reasoning is required; roughly 80% of the workflow is now scripted, deterministic code, with evidence committed to disk before moving on. Intuit faced similar challenges rebuilding its AI agent architecture.
Sponsored Protocol
Walmart: citizen developers and rampant duplication
Walmart hit an unexpected bottleneck: success. An agent harness placed directly into employees' hands went viral internally. Gosby called these "citizen developers" who started building agents to solve problems that once required formal engineering roadmaps. Innovation flourished, but duplication soared: dozens of overlapping agents with no coordination. The solution was not to restrain the harness but to build governance that spots duplication, promotes the best agent version, and pushes it into production without engineering becoming a chokepoint. Brex recently launched CrabTrap, an open-source proxy to govern AI agents using real traffic policies, a related approach.
Sponsored Protocol
Zendesk: data pipeline of 20 billion conversations
Zendesk faced its bottleneck on the data side. Ghoche, who joined through the acquisition of Forethought in March 2026, described a repository of roughly 20 billion customer conversations. The instinct would be to feed that history to a large language model with a big context window, but Ghoche noted that does not work. Instead, companies must invest in underlying data pipelines and data infrastructure to make agents efficient. This highlights how data often becomes the overlooked limiting factor.
The role of open source and vendor independence
All three leaders agreed on one point: own what you can. LinkedIn built an AI gateway that unifies all outbound calls to LLMs regardless of provider, plus a memory subsystem to hold context independent of any model. Walmart created an internal gateway to stay vendor agnostic across three workload types: fully deterministic, planner-and-reasoner, and hybrid. The choice between frontier and open-weight models depends on effectiveness for the specific workload, not a fixed policy.
Sponsored Protocol
Advice for the modernization journey
Ghoche advised investing in evals before anything else, because they force problem decomposition and accelerate progress. Gosby recommended putting an AI harness into employees' hands from day one, paired with monitoring infrastructure. Singh urged building for model and context independence, keeping data within the enterprise to reuse with future models. Read the original article on VentureBeat for executives' full statements. Wikipedia on artificial intelligence provides broader context on this technology.
