Red Hat AI 3.4 adds governance for agentic systems

Red Hat has unveiled Red Hat AI 3.4, a new version of its artificial intelligence platform centred on a single framework for running models and autonomous agents across hybrid cloud systems.

The release targets companies trying to move artificial intelligence projects from pilot stages into day-to-day operations with tighter oversight. It is designed to give developers access to models while allowing infrastructure teams to monitor use, apply policies and track how systems and agents behave.

At the centre of the update is a Model-as-a-Service offering that provides a governed interface for developers to use approved models through standard application interfaces. Administrators can track consumption, set rules for internal and external model access, and use identity-provider-based authentication built into the service.

Red Hat is also adding a set of AgentOps tools to manage AI agents from development through production. These include tracing, observability, identity controls and lifecycle management, reflecting growing concern among businesses over how autonomous software systems make decisions and carry out tasks.

The platform also adds prompt management and an evaluation hub. Prompts can be stored and managed as data assets in a central registry, while evaluation tools assess quality, accuracy, safety and risk across models, applications and agents.

MLflow has been integrated into the platform to provide experiment tracking and artefact management for both generative and predictive artificial intelligence use cases. The integration also gives users visibility into agent execution, including language model calls, reasoning steps, tool execution, responses and token usage.

Operational controls

A key feature of the update is its emphasis on governance and auditability. Organisations need more visibility over autonomous systems because agents can act with a degree of independence, creating security and compliance risks if their actions cannot be traced.

To address that, the platform uses cryptographic identity management based on SPIFFE and SPIRE, replacing static keys with short-lived tokens. This is intended to support least-privilege access and link agent actions to verified identities.

The company is also building in automated safety testing and adversarial scanning. Using technology from Chatterbox Labs and the Garak project, the platform screens models and agentic systems for risks including jailbreaks, prompt injection and bias, while NVIDIA NeMo Guardrails provides run-time safety controls.

Inference remains another focus of the release. Red Hat AI Inference adds request prioritisation so interactive and background traffic can share the same endpoint, with latency-sensitive requests handled first when systems are under pressure. Speculative decoding, now generally available, can improve response speeds by two to three times while reducing the cost per interaction.

Support for the inference stack has been extended beyond Red Hat OpenShift to other Kubernetes services, including CoreWeave and Azure. This gives customers a more consistent inference environment across on-premise systems and cloud deployments.

Partner backing

Joe Fernandes, Vice President and General Manager of Red Hat's AI Business Unit, outlined the company's position on the shift towards autonomous systems.

"The agentic era represents an evolution of our platform from running traditional applications to powering intelligent, autonomous systems," said Joe Fernandes, Vice President and General Manager, AI Business Unit, Red Hat. "We are defining the open standard for how the enterprise executes AI. By providing a hardened, metal-to-agent foundation for AI inference, MaaS and AgentOps, Red Hat provides the operational assurance organisations need to innovate at scale while maintaining rigorous control."

CoreWeave said its work with Red Hat is focused on making the same inference software stack available across cloud and on-premise environments.

"CoreWeave's collaboration with Red Hat is grounded in a shared commitment to openness and delivering a high-performance inference foundation that allows enterprises to scale their most complex AI workloads," said Urvashi Chowdhary, Vice President of Product Management - AI Services at CoreWeave. "Together, we've delivered a deployment blueprint for Red Hat AI Inference on CoreWeave Kubernetes Service to run the same inference stack on-prem and in the cloud, with Kubernetes-native control and production-grade performance. This enables enterprise AI teams in regulated industries to focus on the important work: building and scaling AI, not retooling their stack for every new environment."

NVIDIA also linked the release to a broader need for tighter infrastructure controls around long-running agents in business settings.

"Autonomous, long running agents in the enterprise demand a new level of infrastructure control and security to ensure trustworthy operations at scale," said John Fanelli, Vice President, Enterprise Software, NVIDIA. "Red Hat AI Factory with NVIDIA provides a unified, open source-driven foundation that gives developers and operators the governance and confidence necessary for the agentic future."

Red Hat said the update will support NVIDIA Blackwell GPUs and AMD MI325X hardware from launch, and will also be available through managed cloud environments including IBM Cloud. The aim is to offer a common operating model across different hardware and cloud providers as companies build and run AI systems at larger scale.