Insights from a customer roundtable at Red Hat Summit 2026Platform engineering and operations leaders from across industries—including airlines, utilities, financial services, higher education, and government—gathered for a candid conversation about agentic AI at Red Hat Summit 2026. We wanted to find out what's actually working, where the risks lie, and how teams are finding value today.From platform management to AI collaborationThe question underpinning every conversation at our roundtable was a variant of this: What does it look like when a platform stops being something you manage and starts being something that thinks alongside your team?Teams are managing complex, multicluster environments with the same headcount they had 2 years ago. The manual work of diagnosing incidents, correlating alerts, and executing remediation is consuming capacity that should be directed toward core development. One participant stated their goal plainly—double the number of clusters being managed without hiring a single additional person.The need goes beyond efficiency. The real question is, what do your best people work on when the routine work takes care of itself? The customers at our roundtable are already getting glimpses of it. One team's agents detected a storage driver silently crashing every 15 minutes—invisible to existing logs and alerts. The agent surfaced the issue with a full diagnostic summary and placed itself in a blocked state until a human authorized the next step. That action delivers faster resolution, certainly, but also a fundamentally different relationship between a platform and the people running it.Surprising return on investment (ROI) beyond ITSome of the most compelling insights came when participants described AI agents solving problems outside the traditional infrastructure domain.A university team deployed agents to help their student financial services group identify underserved students who qualified for federal funding but had been overlooked by existing database queries. The system quickly identified more than 100 students and $60,000–$70,000 in unclaimed federal aid. The team hadn't realized this use case was possible a few weeks earlier.Another example was when a government agency described its longer-term vision of providing an agentic interface to help citizens navigate complex legal regulations and automatically coordinate the appropriate public services based on major life events, such as having a child, starting a business, or relocating—giving citizens a single intelligent interface rather than dozens of disconnected portals.The organizations gaining the most traction are those that are expanding their use cases and stretching their preconceived notions of what "AI for operations" can mean.Human oversight is non-negotiableThe roundtable consensus was clear—full autonomy is not an option. Because AI is non-deterministic, the core challenge isn't whether to trust agents, but how to structure workflows so agents can safely accelerate routine tasks while humans retain control over critical actions. Agentic solutions need a clear, standard workflow governance model that’s native to the platform.The model that resonated most was treating agents like site reliability engineers—they analyze systems, generate execution plans, and draft tickets summarizing findings and proposed actions. Engineers review and approve, and the agent only executes what has been explicitly authorized.One team shared a cautionary story from testing supervisory agent architectures, where the AI supervisor agent oversees a set of specialized sub-agents. They shared that 3 subordinate agents coordinated to convince the supervisory AI to approve an action it should have rejected. The team caught it before any damage occurred, but the lesson was clear: AI cannot reliably govern other AI without human intervention.This example points to a product need for codifiable, easy-to-use governance features, such as a "guardrail service" or "Approval-as-a-Service," that organizations can deploy across all environments from Day 1.Safe by designFor operations teams, a major focus is containing the fallout when autonomous tools make mistakes. Risks ranged from an agent attempting a destructive command (like rm -rf) to a diagnostic agent repeatedly running resource-heavy commands and accidentally crashing the very cluster it was deployed to fix.To mitigate these risks, teams are using sandbox agents in containerized environments with strict resource limits and full audit logging for every action. Pairing agents with deterministic data layers, such as schema‑validated document stores or other systems that strictly constrain acceptable inputs, has also proven highly effective at keeping agent behavior grounded.The barriers: Buy-in, audit trails, and ROIThe hurdles slowing adoption are largely organizational rather than technical.Building internal buy-in is difficult, especially among teams who have experienced over-promised automation initiatives in the past. Detailed audit trails are essential to building that trust (and legally mandated in some parts of the world)—organizations must be able to show regulators and leadership exactly what an agent did, the reasoning it used, and who authorized the action. Without that evidence, governance teams default to rejecting the technology.Demonstrating clear ROI is equally essential. Every team at the roundtable noted the pressure to prove value before securing budget. The teams moving fastest are starting with small, high-visibility proofs of concept, documenting results, and using that evidence to gain further funding. A $60,000–$70,000 win in student financial aid is a far more compelling budget conversation than a reduction in alert noise, even if both deliver value.Key findingsOrganizations at this roundtable are at different stages of agentic AI adoption, but they're facing the same questions. A few practices stand out from the teams making the most progress:Start with observability and self-healing: This is where operational pain is highest, the path to value is clearest, and a bounded scope makes the project manageable.Design for human oversight from Day 1: Don't bolt on governance after deployment. Build approval workflows, block states, and audit trails directly into your architecture.Sandbox your agents: Containerized execution environments aren't optional—they provide the necessary guardrails to let agents operate safely.Get serious about AI infrastructure costs: Several teams described GPU and cloud inference costs spiraling faster than expected, in some cases outpacing the value being generated. Red Hat OpenShift AI is designed specifically for this challenge, helping enterprises optimize GPU use across hybrid infrastructure so the cost of running AI at scale doesn't become the reason you can't scale it. In an environment where both cloud inference bills and on-premise GPU hardware represent significant capital commitments, having a platform that makes that investment work harder is crucial.Focus on high-visibility business value: The teams with the most momentum find use cases that deliver value to the entire organization, not just the platform engineering team. Red Hat OpenShift for the agentic eraEvery major wave of enterprise software has required a foundational layer that made complex, distributed systems governable at scale. In the client-server era, it was middleware like JBoss. In the cloud-native era, it was Kubernetes. In the agentic AI era, that role belongs to Red Hat OpenShift and Red Hat OpenShift AI.Customers at this roundtable are building autonomous agents that operate across multicluster environments, execute decisions within strict governance boundaries, integrate with enterprise workflows, and run on sovereign or hybrid infrastructure. This cannot be reliably built on ad hoc tooling. It requires a security-focused, consistent, open source foundation that enterprises already trust. OpenShift is more than an orchestrator for containers. It's the platform on which agent frameworks, MCP servers, agent skill libraries, human-in-the-loop workflows, and GPU workloads are unified into a foundation enterprises can deploy, govern, and scale across multicluster environments.Establishing that security-focused foundation requires addressing the full stack of agentic governance: policy and authority controls, identity and access management, application and tool governance, runtime safety and audit, execution sandboxing, data and content governance, and human-in-the-loop lifecycle management. Critically, identity and authentication cannot be an afterthought—every agent action must be authenticated and attributable to a specific authorized entity. OpenShift delivers on these requirements through AgentOps, Red Hat AI's framework-agnostic approach to operationalizing agents across any model, accelerator, or cloud, with built-in guardrails that address each of these layers of governance and security. These aren't optional features to add on later—they're the foundation that makes enterprise-scale agentic AI possible. The organizations that understand and address this from the beginning will move faster, govern more effectively, and spend less time building this foundation, allowing them to move confidently into the future of agentic AI.To see how Red Hat OpenShift is approaching the future of agentic AI:Watch the Product Spotlight session from Red Hat Summit 2026. Check out our blog, "Models don't generate revenue — applications do." Visit the Red Hat OpenShift AI product page.