DevOps and the Control Problem Behind Agentic AI

The most important DevOps message from IBM Think 2026 was about control, not speed (IBM, 2026). Agentic AI promises faster remediation, smarter operations, and more responsive systems, but those gains depend on whether DevOps teams can govern the identities, permissions, workflows, and production pathways agents will use.

Once AI moves into production, it stops being a model performance story and becomes an operational one. Agents begin interacting with the same systems that DevOps teams already manage every day: deployment pipelines, cloud infrastructure, containers, observability platforms, access controls, security tooling, and incident workflows. The pressure shifts from building AI capability to running AI safely inside live enterprise environments.

DevOps has been the discipline that keeps software moving through the business without losing reliability. It connected developers with operations, automated release cycles, improved incident response, and gave engineering teams better visibility into production. AI now stretches that mandate. The systems being managed are becoming more autonomous, and the actions within those systems are occurring faster than traditional review cycles can comfortably accommodate.

That was the more important message running through IBM Think 2026. Beneath the product announcements was a serious recognition that AI systems require an operating model, not a collection of disconnected tools. That model must include who is allowed to act, the systems an agent can reach, which data can it use, the remediation steps that can happen automatically, where human approval enters the process, and how business can explain at automated decision after the fact.

IBM Senior Vice President for Software Dinesh Nirmal said that generative AI could bring one billion new applications into enterprises over the next five years. Each containerised application may contain hundreds or thousands of microservices.

For DevOps leaders, that means more releases, more dependencies, more identities, more policy decisions, more telemetry, and more operational risk moving through environments that are already difficult to govern.

A DevOps team can have strong monitoring, solid CI/CD, reliable infrastructure automation, and mature incident response, yet still struggle when agents begin acting across those layers. An agent may recommend a remediation, trigger a workflow, request privileged access, or change infrastructure based on a signal from another system. At that point, DevOps needs context, policy, ownership, and a clear record of action:

“For every human access in an enterprise, there is going to be 120 non-humans that will be accessing the enterprise,” said Dinesh Nirmal, SVP for software.

Agentic systems multiply machine identities, and every identity needs scope, policy, and auditability. DevOps teams have long worked with secrets, permissions, deployment credentials, and infrastructure access. AI increases the importance of that work because agents may operate continuously across environments. Concert Protect and Secure Coder point toward a model where exposure management moves closer to the developer workflow, allowing risk to be detected and addressed where code, infrastructure, and policy decisions are being made.

AI agents introduce a new operating pressure because they can participate in the same workflows DevOps teams run every day. An agent may analyse telemetry, suggest a rollback, trigger a deployment task, open a remediation workflow, adjust infrastructure, or recommend a configuration change. Each action depends on access. Each access point carries risk. Each automated step needs to be traceable.

That makes DevOps the natural control layer for agentic systems. Security teams can define policy, engineering teams can build applications, but DevOps is where those policies and applications meet production reality. It is where permission models, automation rules, runtime behaviour, and system reliability have to work together.

Automation also needs a harder look. Pipelines, runbooks, infrastructure templates, and incident workflows were originally designed for human teams. Agentic systems can use those same pathways at far greater speed. Weak workflows, unclear approvals, or loose permissions can scale problems quickly.

DevOps now has to provide the structure that allows AI to act productively inside the enterprise. That means policy aware automation, connected telemetry, real time infrastructure context, secure machine identities, and a clear chain of responsibility when systems make or recommend decisions.

AI agents become production actors as they will build controls into CI/CD, infrastructure as code, secrets management, access governance, observability, and incident response. They will also make agent activity readable after the fact: which workflow ran, which system changed, which permission was used, and whether the outcome stayed within policy.