AI governance shifts from observability to provable control as agents enter production

The report said the challenge is becoming more urgent as agentic systems spread across regulated enterprises, where security, auditability and sovereignty requirements can limit reliance on external software-as-a-service control planes. In the latest episode of theCUBE Research’s AppDevANGLE podcast, Sudeep Goswami, chief executive officer of Traefik Labs Inc., and Andreas Prins, who leads sovereignty strategy at SUSE Group, discussed how AI governance architectures must evolve.

Goswami said authority should narrow when an agent hands a task to another agent. “When you have an agent that is handing a task to another agent, that authority should shrink and not leak out,” he said.

Traditional enterprise observability has focused on reconstructing what happened after an event. Logs, traces and dashboards can show which systems were involved and where a failure occurred. Agentic AI introduces a different governance problem, the report said. An autonomous system may receive authority from a human, delegate a task to another agent, invoke a tool or application programming interface and then pass part of that task to another system. Every step can change the scope of what is authorized.

Identity and credentials alone are insufficient in that environment. “Just because an agent has some credentials, is that agent allowed to make this specific action right now, given the surrounding context around it?” Goswami asked. “A simple credential cannot answer that.”

The report compared the problem to physical access control. An employee badge may authorize entry into a building, but it does not automatically grant permission to approve a financial transaction or access every system inside the organization. AI governance therefore needs to become contextual, with policies determining what an agent can do based not only on its identity but also on the task, environment, delegation chain and operational conditions surrounding the request. That shift moves governance from post-event reconstruction toward real-time authorization and proof.

Agent delegation adds further complexity. Traditional enterprise access models were designed largely around humans, applications and predefined service accounts. Agentic systems introduce machine-to-machine interactions in which authority can move across orchestrators, subagents, APIs and tools at machine speed.

Prins compared the shift to the evolution of continuous integration and continuous delivery systems. Years ago, software teams moved from manual production releases toward automated pipelines. That required organizations to codify approvals, security checks, secrets management and release policies directly into the delivery process. “We go through that same thinking again,” Prins said. “Let’s rethink, and more importantly, let’s articulate as code.”

The issue grows as the number of enterprise agents increases. Prins described a recent conversation in which an executive discovered that an engineering organization had created approximately 8,000 agents. “If you’re unaware how many are created, you are also unaware what they’re doing and what their function is,” he said.

The report said agent proliferation could quickly outpace traditional governance models. Enterprises may eventually need the equivalent of software supply chain controls for agents: explicit identity, delegated authority, policy enforcement and evidence showing how each system behaved.

Defining policies is only part of the problem. Enterprises also need mechanisms capable of enforcing those policies where agents interact with applications and infrastructure. That puts the gateway layer in an increasingly important position, according to the report.

Goswami argued that AI governance systems need to capture both allowed and denied actions. Showing that a system successfully blocked an unauthorized action can be as important as showing that an authorized task completed correctly. “You want to be able to showcase proof that your guardrails are working,” he said. “Which means it should allow and deny, and you want to have that entire thing in context.”

The report said the enforcement layer can also provide a feedback loop into governance policies themselves.