Beyond Copilots: What AI Agents Need to Understand Cloud Operations

Why useful infrastructure investigations depend on operational context, clear boundaries, and verifiable evidence.

Giving an AI agent access to infrastructure tools creates possibilities. It also raises an engineering question: how can that agent investigate a real cloud environment while respecting permissions, operational boundaries, and security requirements?

This question shapes our perspective at Opsphere.

As José Julián Ariza, CEO of Opsphere, joins the AI Congress Barcelona session on AI agents, multiagent systems, and orchestration, we want to share the practical challenge behind that discussion.

Engineers need more than an agent that can call tools. They need help understanding what is happening in production, why it might be happening, and what evidence supports the next decision.

Consider an engineer investigating a slow application.

An agent might be able to retrieve logs, inspect infrastructure, and review a recent deployment. Those capabilities provide useful observations.

Turning them into an investigation requires knowing how the observations relate.

Which environment serves the affected application? Does the deployment belong to the service under investigation? Are the errors recent? Is the apparent dependency actually part of the affected request path?

Without those relationships, an agent can collect accurate information and still reach an incorrect conclusion.

Operational understanding requires both reliable observations and enough context to interpret them.

Production infrastructure contains distinctions that matter: separate accounts, environments, services, teams, and permission scopes.

An investigation must preserve those boundaries.

An engineer asking about a production endpoint should receive findings grounded in that environment. Evidence from a similarly named service in another account could lead the investigation in the wrong direction.

Permissions also shape what an agent can establish. If access to a relevant system is unavailable, that limitation should remain visible. Missing evidence should become an explicit gap in the investigation.

The distinction between observation and action matters equally. Permission to inspect a resource does not automatically authorize changing it. A workflow that retrieves logs has different consequences from one that modifies configuration or restarts a service.

Useful agents need a clear scope for both the information they can access and the operations they can perform.

Suppose errors increase shortly after a deployment.

The timing makes the release worth investigating. It does not, by itself, establish the cause.

The agent should examine whether the affected requests reached the new version, whether the changed code could explain the symptoms, and whether a dependency was already degrading.

A useful investigation makes its reasoning inspectable:

These questions turn a plausible narrative into something an engineer can assess.

The distinction becomes especially important under pressure. A confident explanation can influence a production decision even when its supporting evidence is weak.

Engineers need to understand how a conclusion was reached and what remains unverified.

Multiple agents can divide investigative work. One might examine infrastructure state, another review deployment history, and another inspect application errors.

That division is useful when each contribution remains tied to the same incident scope.

The agents need to work against compatible environments and time windows. Their findings need identifiable sources. Contradictions need to survive the handoff so the overall investigation can address them.

Agreement alone is insufficient. Several agents may repeat the same assumption or rely on the same incomplete source.

Orchestration therefore involves more than assigning tasks. It must preserve the relationships between questions, evidence, findings, and unresolved uncertainty.

The result should help an engineer follow the investigation without reconstructing every step.

At Opsphere, our direction is grounded in this operational challenge: helping engineers use AI to understand complex cloud environments and make better-informed decisions.

That means treating context, access boundaries, and evidence as central parts of the engineering problem.

The questions guiding this work are practical. Can the investigation identify the right environment? Can findings be traced to operational information? Can the engineer distinguish an observation from an inference? Does the workflow make its limitations clear?

These are useful criteria for evaluating AI agents in cloud operations, alongside the tools they can access and the tasks they can coordinate.

Moving beyond copilots should produce a concrete improvement in how engineers investigate systems.

An engineer should be able to ask a scoped operational question, inspect the supporting evidence, understand the uncertainty, and decide what to do next.

That is the direction we are working toward at Opsphere—and the perspective we bring to the conversation at AI Congress Barcelona.

More context. Less friction. Better engineering decisions.