The wall
The model conversation is the easy one. Orchestration and security are where programs stall. Several agents, several tools, and a pile of third-party APIs, with nothing in the middle that knows the caller.
Shared keys
The pattern is familiar. One key in an env file. Copied into a second service because it was faster. No identity on the request. No owner when usage spikes overnight. It is a fine prototype. Production asks who made the call, which team owns it, and who pays.
With a shared credential, nobody can answer.
What Kimss takes, and what it does not
Kimss does not take over your infrastructure, and it does not host the model. You keep the endpoint. You point the client at the gateway, and the call picks up a workspace identity. Model keys stay in your vault. Kimss runs the gate.
Guardrails are a separate decision from “we have a gateway.” Content safety and Prompt Shields are the workspace baseline. Stripping secrets and personal data before they leave is Production and above. Web search stays off until someone turns it on and the consent is recorded.
Developers keep the client they already write. Platform and security get a place to see the routed call, cap it, and turn that agent off at the gateway if it should stop.
Adapted from a LinkedIn note on 16 June 2026.