Everyone building agentic systems is focused on capability: what can the agent do, how far can it reason, how reliably does it complete tasks? These are the right questions for a research agenda. They are not quite the right questions for an enterprise deployment.
The more interesting question is: what happens when the agent tries to do something that requires a change request?
In most large enterprises, "make a change to a production system" is not a single action. It is a process — sometimes a long one — involving approvals, documentation, risk assessment, and a CAB meeting that meets on Thursdays. An agent that can execute end-to-end but hits a wall at the governance layer isn't a productivity multiplier. It's a new source of queue congestion.
The agentic AI that actually sticks in enterprise environments will be the one that understands organizational topology — not just system topology. It will know not just how to connect to a system, but who needs to sign off before it touches one. That's a harder problem than reasoning, and it's the one I find most interesting right now.