Red Hat adds agent governance

Red Hat on May 12 rolled out Red Hat AI 3.4 with new governance, observability and safety features aimed at companies trying to move AI agents from pilots into production. The release, announced at Red Hat Summit in Atlanta, packages model serving, policy controls, evaluation tools and agent operations into what the company calls a “metal-to-agent” platform. Red Hat said the update is built for hybrid-cloud deployments, where enterprises want to run models and autonomous workflows across on-premises and cloud infrastructure. The company framed the release around a practical problem: once agents begin acting across tools and data sources, operators need more than model access — they need controls over identity, usage, safety and lifecycle management. ### Which new controls did Red Hat actually add? Red Hat said Red Hat AI 3.4 introduces Models-as-a-Service, or MaaS, as a governed interface for developers to access approved models while administrators track consumption and enforce policies. In the company’s description, the feature includes self-service token key management for role-based administration, support for self-hosted models and some cloud-based models, and usage tracking and showback capabilities. (redhat.com) The May 14 OpenShift AI 3.4 post gave more detail on those controls. Red Hat said administrators can set token quotas and rate limits through Kubernetes-native custom resources, while developers generate API keys tied to specific subscriptions and those keys can be revoked. The company also said showback dashboards in technical preview track token consumption by model and subscription group. (redhat.com) ### How is this different from ordinary model hosting? Red Hat said the release is designed not only to serve models but also to manage the behavior of agents that call tools, APIs and enterprise systems. Its agentic AI product page says customers can control how agents behave, what they can access and how they interact with data and tools in production, using guardrails, identity and human-in-the-loop checkpoints. (redhat.com) The company’s press release said newly introduced AgentOps tools cover tracing, observability, cryptographic identity and lifecycle management from development through production. Red Hat also said the platform is framework-agnostic, allowing customers to deploy agents built with different software stacks rather than forcing a single approach. (redhat.com) ### Where do safety testing and audit trails come in? Red Hat said Red Hat AI 3.4 adds an “evaluation hub” for testing model and agent accuracy, quality and safety. A Red Hat Developer post published May 12 described EvalHub as a unified control plane based on the upstream EvalHub project, built to orchestrate evaluation jobs across Kubernetes or Red Hat OpenShift clusters. (redhat.com) The same post said the hub can work with tools such as Garak and lm-evaluation-harness, or with proprietary scripts, and records results as MLflow experiments while generating immutable OCI artifacts for an audit trail. Red Hat’s press release also said the platform uses technology from Chatterbox Labs and the Garak project for automated safety testing and red-teaming of models and agents. (redhat.com) ### Why is Red Hat talking about builders and operators together? Red Hat’s May 12 announcement said the handoff from AI experimentation to production requires a common framework for developers and infrastructure teams. The company said Red Hat AI 3.4 is meant to give builders model access and deployment tools while giving operators policy enforcement, security controls and infrastructure efficiency. (developers.redhat.com) Jennifer Vargas, Carlos Condado, Younes Ben Brahim and Will McGrath wrote in Red Hat’s May 12 blog post that autonomous agents are “resource intensive by design,” operating continuously and hitting infrastructure many times to complete a task. They said that combination of sustained compute demand, cost pressure and governance requirements can block production deployments without a dedicated operating foundation. (redhat.com) ### What comes next in this rollout? Red Hat’s published materials show that not every part of 3.4 is generally available yet. The company said some MaaS capabilities, including showback, are in technical preview, and Red Hat documentation for related 3.4 inference components also labels some builds as early access rather than production-ready. (redhat.com) Red Hat said customers can review the Red Hat AI product pages, the May 12 Summit announcement and the 3.4 documentation on its website for release status and feature availability. The next visible step is broader rollout of the preview and early-access features into supported production releases across the Red Hat AI portfolio. (redhat.com 1) (redhat.com 2)