Agentic
GenAI, LLMs, RAG, and autonomous agents as the cross-cutting method-revolution that multiplies infrastructure, operations, and the software lifecycle.
This is the newest axis, and the one that behaves differently from the other three. It is not a fourth sibling to infrastructure, operations, and SDLC — it is a method-revolution applied across all three. It changes the way the work is done, adding velocity and autonomy; it does not change the fundamentals underneath.
At its core it is hands-on engineering of the model-and-agent machinery: frontier and open LLMs and their internals (transformers, mixture-of-experts, diffusion), RAG grounding over private knowledge, composition into autonomous tool-using agents (MCP, A2A, orchestration frameworks, coding-agent skills), and the physical-AI and world-model frontier (Omniverse, OpenUSD, computer vision).
Unlike the other domains, this one is breadth rather than depth. It is measured in a handful of years, not decades — which is exactly why the durable disciplines from the other three are what keep it honest.
The cross-cut
The same fundamentals, at higher velocity and greater autonomy. What follows is the same revolution seen from each of the other three domains.
In infrastructure
An engineer designs and stress-tests hardware and network topology as a world model before any metal is racked. An accelerated-computing site — Blackwell, Grace, BlueField — and its GPU fabric are built inside an OpenUSD scene in NVIDIA Omniverse, so that a digital twin of the AI factory validates power, cooling, and cabling in simulation.
The same agent then generates and validates the infrastructure-as-code that instantiates it: InfiniBand and Spectrum-X fabric configuration, RDMA and PKey setup, NCCL topology, Kubernetes GPU-scheduling manifests — checked against congestion and lossless-Ethernet constraints. This catches misconfigurations that once took a lab and several weeks to surface.
In operations
An engineer investigates and triages incidents by pointing an LLM-backed agent at the logs, metrics, and traces flowing through OpenTelemetry and Prometheus, correlating telemetry into root-caused findings that previously took manual dashboard-spelunking. The fix is then driven through MCP servers that expose Ansible playbooks and Kubernetes control planes as callable tools, so remediation happens inside the same conversational loop — the pattern behind natural-language Day-2 fronts such as OpenShift Lightspeed.
The net effect pushes the collect→correlate→remediate loop past human-in-the-loop dashboards and toward agentic remediation, where the agent both diagnoses and, under policy, closes the loop. This is the oldest control loop in operations finally getting the reasoning step it always implied.
In software development
An engineer drives the build-and-deliver loop through a coding agent — describing intent in natural language and letting the agent write and refactor code, generate tests and configuration as code, and run git itself to stage, branch, and commit — while supervising: prompting, reviewing diffs, steering across multi-step tasks.
The durable discipline does not disappear. It inverts. Version control, CI/CD, and code review become the guardrail around agent-authored code: the deterministic gate that verifies what the non-deterministic agent produced, before it ships. That inversion is the single most important thing to understand about this domain — the old rigour is not obsolete, it is load-bearing in a new way.
Why the other three still matter
An agent that generates fabric configuration is only as good as the constraints it is checked against, and those constraints come from infrastructure. An agent that closes a remediation loop is safe only inside the policy envelope that operations defines. An agent that writes code ships safely only through the gate that SDLC built.
Agent failure, in practice, is control-loop failure. The discipline that makes agents useful is the discipline the other three domains have been accumulating for thirty years.
Agentic transformation sits on top of the other three domains rather than beside them. Start with infrastructure for the substrate, operations for the loop, and SDLC for the gate.