Beyond Copilots: What AI Agents Need to Understand Cloud Operations

Investigations in infrastructure depend on context, defined limits, and proof. Allowing an AI agent access to infrastructure tools opens opportunities, yet also raises a technical question: how can that agent examine a live cloud setting while abiding by permissions, operational boundaries, and security standards? Opsphere, led by CEO José Julián Ariza, is addressing this practical issue at an AI Congress Barcelona session on AI agents, multiagent systems, and orchestration.

Engineers require more than a tool-accessing agent. They need assistance understanding what is occurring in production, why it might be happening, and what evidence supports the following choices. Gaining access to tools marks the initial step. However, turning those capabilities into an investigation necessitates grasping the relationships between observations.

Which environment does the troubled application rely on? Does the deployment associated with the investigation belong to the affected service? Are the errors recent? Is the perceived dependency genuinely part of the request's pathway? Without these connections, the agent could gather precise data yet still draw the wrong conclusion.

Operational comprehension demands both dependable observations and sufficient context for interpretation. Live environments have distinct boundaries: separate accounts, environments, services, teams, and permission scopes. When investigating a production endpoint, an engineer should receive conclusions rooted in that environment.

Data from a similarly named service in another account could mislead the investigation. Permissions also influence what an agent can establish. If access to a relevant system is unavailable, that limitation should remain visible. The discrepancy between observation and action is equally significant. Permission to observe a resource does not automatically permit altering it.

A workflow that retrieves logs has different consequences than one that changes configuration or restarts a service. An explanation requires corroborating evidence. Suppose errors escalate following a deployment. The timing indicates the release is worthy of investigation. Nevertheless, this evidence alone does not establish the cause.

The agent should investigate whether the impacted requests reached the new version, whether the altered code could explain the symptoms, and whether a dependency was already deteriorating. A valuable investigation should make its reasoning transparent: What has been directly observed? Which prediction does this observation support?

What evidence refutes the prediction? What information is still missing? What check would assist in resolving the uncertainty? These questions transform a plausible narrative into something an engineer can evaluate. This distinction becomes particularly critical under pressure. A self-assured explanation can affect a production decision even when its supporting evidence is weak.

Engineers must comprehend how a conclusion was reached and what remains unverified. Multiagent systems require a shared investigative approach. Multiple agents can divide investigative tasks. One might analyze infrastructure state, another review deployment history, and another examine application errors. This division proves helpful when each contribution is linked to the same incident scope.

The agents must operate within consistent environments and time frames. Their findings must bear identifiable sources. Discrepancies must hold up to the handoff so the overall investigation can address them. Consensus is insufficient. Multiple agents may repeat the same assumption or depend on the same incomplete source. Orchestration entails more than allocating tasks.

It must maintain the relationships between questions, evidence, findings, and outstanding uncertainties. The desired outcome should assist an engineer in following the investigation without reassembling every step. Opsphere's objective is to help engineers utilize AI to grasp intricate cloud environments and make more informed choices.

This goal entails treating context, access boundaries, and evidence as fundamental elements of the engineering challenge. The guiding questions for this work are practical: can the investigation pinpoint the appropriate environment? Can the conclusions be traced to operational information? Can the engineer distinguish an observation from an inference?

Does the workflow make its limitations evident? These criteria serve as useful benchmarks for assessing AI agents in cloud operations, along with the tools they can access and the duties they can coordinate. Transitioning beyond copilots should result in a tangible enhancement in how engineers investigate systems. An engineer ought to be able to pose a scoped operational question, review the supporting evidence, grasp the uncertainty, and decide the following steps.

That is the direction Opsphere is pursuing—and the viewpoint we contribute to the discussion at AI Congress Barcelona. Greater context. Fewer obstacles. Better engineering choices.