LinkedIn, Walmart and Zendesk point AI agent delays to old infrastructure

LinkedIn, Walmart and Zendesk executives said at VB Transform 2026 that the hard part of putting AI agents into production has been rebuilding surrounding infrastructure, rather than swapping in better models. The panel did not disclose spending, deployment counts or latency targets, but the examples showed where enterprise AI work is shifting: orchestration, governance, data systems and model abstraction.

The discussion included 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 a different failure point once agents moved beyond pilots.

The common issue was that internal systems had been designed around human-paced work. Agents create a different operating profile, with repeated calls, fast handoffs and less tolerance for slow provisioning or manual coordination.

Production agents exposed old assumptions

At LinkedIn, Singh said Kubernetes became an early constraint because container startup times are measured in seconds. That was too slow for the company’s agent workloads, so LinkedIn moved to pre-provisioned container pools that can take on and release agent tasks in real time.

LinkedIn also changed how it lets agents manage workflows. Singh said a five-point evaluation approach still let hallucinations appear, because using one LLM to judge another LLM can reproduce the same error pattern. The company responded by limiting where LLMs sit in the process.

“We built our own harness, our own control flow, and pushed the LLMs to the leaf instead of them orchestrating the loop,” Singh said. He said about 80% of the workflow now runs as deterministic code, with LLMs reserved for reasoning steps. Evidence from each step is written to disk before the system proceeds.

Walmart ran into a different scaling problem after an internal agent harness spread among employees. Gosby said “citizen developers” began creating agents for tasks that might previously have waited for engineering resources. That produced useful internal experimentation, but also overlapping tools built for similar purposes.

Walmart’s response, according to Gosby, was governance that can detect duplication, select the best version of an agent and move it into production without turning engineering into the approval bottleneck. She said the goal is to ensure “engineering doesn’t once again become the bottleneck for what it is we’re trying to do.”

Zendesk’s constraint was data infrastructure. Ghoche joined Zendesk through its acquisition of Forethought, which closed in March 2026. He pointed to what he described as a public figure of 20 billion customer conversations in Zendesk’s repository, and said that history cannot be handed wholesale to a large language model with a large context window.

“You can’t really do that, so instead you have to really invest in the underlying data pipelines and all the data infrastructure that comes with that,” Ghoche said.

Model independence is becoming a design requirement

The executives also described efforts to avoid dependence on a single model provider. Ghoche said most enterprises would prefer to own models and infrastructure where they can, while still using frontier labs for reasoning work where those providers retain an advantage. He said that category is narrowing relative to the broader range of enterprise AI work.

LinkedIn built an AI gateway so outbound model calls use the same interface across public cloud systems and the company’s own data centers. Singh said the setup lets LinkedIn switch providers more quickly. The company also built a memory subsystem intended to keep enterprise context separate from any individual model vendor.

Walmart has built an internal gateway for deterministic workflows, planner-and-reasoner workflows and hybrid workloads. Gosby said compliance-sensitive tasks remain deterministic by design, while governance, security and evaluation run through the gateway no matter which model is used. She said Walmart chooses between frontier and open-weight models based on the workload.

Evaluation comes first

Ghoche said evaluation should be treated as a starting point across internal and customer-facing use cases. “The thing that’s common to all of these is evals. It’ll force you to break the problem down, and once you have a robust set of evals, you can move a lot faster,” he said.

Gosby argued for giving employees access to the agent harness early, paired with monitoring of what they create. Singh said companies should design for both model independence and context independence, so enterprise context can be reused as models and harnesses change.

This story draws on original reporting from VentureBeat.