Deploying an agent does not make platform engineering obsolete. VMware Tanzu’s Nick Kuhn argues on Practical AI that enterprise workloads still need approved dependencies, restricted networks, durable state, clear identity, isolation, patching, and observable operations; a laptop demo that can fetch anything from the internet is not the same environment. He describes applying a Cloud Foundry-style buildpack and service-binding workflow to agents, with approved models, MCP tools, gateways, and project memory attached through a repeatable path to production. The transferable lesson is the separation of concerns: instructions describe intended behavior, permissions determine actual authority, persistent services preserve state across disposable instances, and telemetry records what happened. An MCP gateway can narrow tool exposure and help preserve the user’s identity instead of routing every action through one generic account, but neither a gateway nor a container proves that every decision is safe. Kuhn explicitly avoids claiming that basic controls prevent all agent incidents. His adoption advice is similarly grounded: start with one useful workflow, add capabilities as needed, and involve existing platform and security teams rather than creating an isolated AI organization. The closing anecdote is a useful warning about the economics of attention: he is working harder because his agents need things from him. More generated work is not automatically less work to review, coordinate, and maintain.
“They want the same repeatable pattern across the board so that they can audit it and secure it.” — Nick Kuhn, 08:16
“You have to have the individual identity pass through.” — Nick Kuhn, 25:11