Cattle, not pets: why agentic AI forces a rethink of infrastructure on Azure

The article, published by PwC as part of its alliance with Microsoft, revisits a classic cloud infrastructure metaphor — treating servers as replaceable 'cattle' rather than as 'pets' cared for individually — to argue that the arrival of agentic AI turns that old, long-ignored recommendation into an urgent necessity. For years, according to the author, many organizations adopted the vocabulary of 'cattle' but retained the emotional attachment and informal practices of 'pets': virtual machines with proper names that nobody dares touch on a Friday, one-off configurations that exist only in one engineer's memory, Infrastructure-as-Code repositories abandoned as soon as someone 'quickly fixes something' manually from the portal, Entra permissions that outlive their original purpose, or approvals that happen in a Teams chat instead of in an auditable pipeline.

Those practices were costly but tolerable when the environment was operated by a person with tribal context. The problem, the text argues, is that agentic systems — Copilot agents, Microsoft Foundry flows, autonomous pipelines — have no way of 'knowing' that a server called SQLPROD03 is special, or that 'the network team always does it this way.' An agent reads policies and available capacity and acts accordingly; if the environment depends on undocumented exceptions, the agent may fail against them or behave in ways the organization did not anticipate.

Hence the article's central proposal: 'agent-ready infrastructure' means environments provisioned, governed and decommissioned through code, with identity, observability and cost control built in by design, not bolted on afterward. This turns what was once a purely operational conversation (automation, orchestration, elastic capacity) into a governance conversation that demands an explicit operating model.

To structure that model, the article proposes three levels of autonomy defined by risk rather than by task type: pre-authorized autonomy (scaling within already approved limits, restarting a failed service, correcting a known drift from an Azure Policy baseline); supervised autonomy (provisioning new resource types or modifying network topology, but passing through an approval gate in Azure DevOps or GitHub Advanced Security, with an audit trail in Microsoft Purview and time-bound access via Entra Privileged Identity Management); and approval-gated autonomy (any action touching identity boundaries, production data planes or regulated workloads, where a human must approve before execution, not after reading the incident report). The key idea is that the level of autonomy must be codified in policy before the agent acts, not decided on the fly, and that it must also account for operational context (a routine action on an ordinary day can be critical during an incident or a peak traffic window).

As a concrete example of this direction, the article cites Microsoft's Azure SRE Agent, now generally available, capable of investigating production incidents, proposing mitigations and automating responses within guardrails and approval flows defined by each organization. It presents this as a signal of where operations on Azure are heading: agents working within defined controls and configurable autonomy levels, not a wholesale replacement of human oversight.

The text also qualifies the usual argument about 'moving beyond virtual machines' toward cloud-native architectures. It acknowledges that VMs are not going away — they remain appropriate for OS-level dependencies, compliance-bound workloads or systems not yet ready for cloud-native — and clarifies that the debate is not 'VMs or no VMs,' but 'pets versus cattle' applied to the entire technology estate: the VM, the container, the serverless function, the identity, the network route and the policy that governs them together. A cloud-native workload migrated by hand, with undocumented exceptions and modified by whoever is on call, is still a 'pet,' just a fashionable one. What agentic AI adds to that argument is a new requirement: the environment must be machine-readable, and an environment governed by explicit policies is more readable than one built on manual exceptions.

In practical terms, the article lists the concrete work needed to reach that 'agent-ready' state: Infrastructure-as-Code coverage across the entire relevant estate (including the exceptions currently handled by hand); scoped, expiring agent identities through Microsoft Entra, so that every action is attributable to a specific, revocable identity rather than to shared credentials; policy as code spanning compute, identity, network and security together in Azure Policy and Entra; observability and auditing that make it possible to reconstruct what an agent did, what it accessed and why; defined rollback paths for autonomous actions, proportional to their risk and impact; FinOps controls that cap how much an autonomous scaling event can spend, not just what it can touch; and, again, a tiered autonomy model, defined in advance and enforced through Entra PIM, approval gates in Azure DevOps and audit trails in Purview.

The article closes by placing this discussion within a larger challenge: it cites a PwC survey in which 45% of the CEOs polled do not believe their companies will still be viable in ten years if they continue on their current trajectory. The author's conclusion is that the opportunity is not autonomy for its own sake, but building an operational foundation that allows more repetitive work to be delegated to AI within defined controls, freeing experienced engineers for higher-value problems. The constraint, he suggests, now lies less in the capabilities of AI agents than in whether the environment around them is ready for them to operate within clear boundaries of policy, access and human oversight.

It is worth noting that this is PwC content tied to its technology alliance with Microsoft and aimed at promoting Azure infrastructure assessment and modernization services, so it combines a reasonable technical diagnosis with an explicit commercial purpose (it includes a call to purchase a PwC 'Azure readiness assessment'). Even so, the underlying diagnosis — that technical debt and undocumented exceptions, tolerable with human operators, become failure points when autonomous agents act — is consistent with broader concerns the industry is raising about AI agent governance in enterprise environments.