Unity expands AI gateway governance

Databricks expanded Unity AI Gateway on May 19 with new beta features aimed at the runtime layer of AI systems, adding LLM guardrails, payload logging for MCP traffic, service policies and cost controls. The company said the release is meant to govern “every model call, tool invocation, and agent interaction” from one layer inside Unity Catalog. Microsoft’s Azure Databricks documentation describes Unity AI Gateway as a beta “central AI governance layer” for agents, LLM endpoints, MCP servers and coding agents. What changed is less about model access than about operating controls. Databricks said guardrails can now be applied to model inputs, outputs or both, while payload logging captures full request and response data for MCP interactions. The same release adds service policies to control which tools agents can invoke based on identity and request context, plus cost controls and per-user alerts across models and providers. (databricks.com) ### Why does payload logging matter more than another model connector? Databricks said payload logging is designed to give teams “complete observability” into MCP traffic, with centralized logging into Unity Catalog. Microsoft’s documentation says Unity AI Gateway can log LLM traffic across endpoints, audit requests and responses in Unity Catalog Delta tables, and monitor usage and costs with system tables and dashboards. (databricks.com) That makes the gateway look more like an operations layer than a model proxy. MLflow’s Databricks integration page describes the same stack in terms of governance, monitoring, payload logging, safety controls and reliability features such as fallbacks and load balancing. ### How does this fit with what Vercel is shipping? Vercel has been making a parallel pitch around AI Gateway as the place where developers unify model access, usage tracking, cost tracking, retries and failover. (databricks.com) Its AI Gateway pages say teams can route to hundreds of models through a centralized interface, and its recent production report says the service has handled tens of trillions of tokens from more than 200,000 teams. Google’s Gemini 3.5 Flash is part of that push. (mlflow.org) Google introduced Gemini 3.5 Flash on May 19 as a model built for agents and coding, and said it is available to developers through Google’s own platforms. Vercel’s model catalog separately lists Google models on AI Gateway, showing how gateway vendors are using fresh model launches as distribution points inside their own control planes. ### Where does Upstash fit into the picture? (vercel.com) Upstash is addressing a different part of the stack: persistent execution. Its Box documentation says each Box is an isolated cloud container with a filesystem, shell, network stack and runtime, with state that persists across runs and optional keep-alive behavior. Upstash’s setup guides for OpenClaw and Hermes show the practical features attached to that model. The OpenClaw guide walks users through SSH access, gateway startup, dashboard tunneling and auto-restart. (blog.google) The Hermes guide recommends a Medium Box and the same SSH-based setup path for a longer-running agent environment. Those details match the broader market move toward durable agent runtimes with logs, shells and restartable state, rather than one-shot API calls. (upstash.com) ### What are gateway vendors converging on? The common feature set now centers on governance, observability and runtime durability. Databricks is adding guardrails, policy controls and payload logging. Vercel is emphasizing unified routing, usage and cost tracking, retries and failover. Upstash is packaging persistent containers for agent frameworks with SSH access and stateful execution. (upstash.com) Databricks’ own framing is explicit. Its May 19 post says teams are facing “rising costs, unclear agent behavior, and limited control” as agents move into production. The next step for users is concrete: Microsoft’s documentation says account admins can enable Unity AI Gateway beta access in the account console previews page, while Vercel and Upstash continue to publish model and setup documentation for gateway and runtime deployments. (databricks.com)