Agents think in milliseconds, legacy infrastructure doesn't. LinkedIn, Walmart and Zendesk shared how they…

**Agents** reason in milliseconds. **Legacy infrastructure** does not. That gap — not model quality — is what three infrastructure leaders from **LinkedIn**, **Walmart**, and **Zendesk** said is actually slowing **AI agents** in production, speaking at **VB Transform 2026**.

The panel put the bottleneck on systems around the model: platforms, corporate technology services, and the stacks that still serve request/response workloads agents outrun. **Animesh Singh**, senior director of AI platform and infrastructure at LinkedIn, and **Desiree Gosby**, SVP of corporate technology services and technology strategy at Walmart, framed the problem as an infrastructure and platform design issue rather than a pure model-scaling one. **Zendesk** joined that same diagnosis from its own stack.

For engineers building agent systems, the constraint is not “better prompts” or a newer model alone. Agents that plan and act in milliseconds still sit on data paths, auth, tooling, and service boundaries built for slower, human-paced or batch-oriented flows. If those layers stay chatty, sequential, or tightly coupled to legacy services, end-to-end latency stays high even when the model is fast.

That puts the competitive edge less on who has access to a frontier model and more on who can rewire platforms so agent loops do not stall on internal systems. LinkedIn’s AI platform work, Walmart’s corporate technology and strategy layer, and Zendesk’s customer-facing stack all face the same class of problem: production agents that think quickly still depend on infrastructure that was not designed for that tempo.

The practical takeaway is to measure agent performance at the full path — model plus tools, data, and legacy services — not at model tokens alone. Watch for concrete platform moves that cut handoffs and idle time around the model: tighter integration of agent runtimes with internal APIs, fewer serial dependencies, and infrastructure owned as an agent product surface rather than a fixed backend. Until those layers match agent speed, model upgrades will keep under-delivering in production.