What Running Agents at Scale Actually Teaches You About Shortage

Here is the pattern I keep running into. Every time AI removes a bottleneck, it does not remove the work, it moves the work somewhere less visible. Enterprises are about to push agents into production at scale over the next two to three years, and the shortage that follows will not be where the headlines are pointing. It will not be "AI engineers." It will show up in storage, in security, and in physical infrastructure, three unglamorous layers that do not get a keynote slide.

Interfaces will probably get easy. Securing them will not.

Start with the part everyone assumes is the hard problem. Most enterprise software still runs on a human interface. Someone reads a screen, clicks a button, follows a runbook step by step. Prompt engineering exists as a job today for exactly one reason. Most systems were not built for machines to operate, so we ask AI to operate them by pretending to be a human with a keyboard. ServiceNow $NOW tickets, Salesforce $CRM workflows, SAP approval chains, Twilio $TWLO, Workday $WDAY, Oracle $ORCL, almost all of it assumes a person is on the other end.

Here is my honest read on this, and it is different from the popular one. I do not think rebuilding these into agent-native interfaces stays hard for long. AI is good at exactly this kind of translation work, reading a legacy API, wrapping it, generating the connective code. Google's API surface and documentation already show what "built for machines" looks like, and I would bet AI-assisted tooling closes that gap across the rest of enterprise software faster than most people expect. Cloudflare $NET already shipped a working example of this. In June 2026 they released an open source library of agent skills that gives an AI agent the vendor-translation logic to migrate a company off Zscaler $ZS or Palo Alto Networks $PANW into Cloudflare's own Zero Trust stack, the same migration logic Cloudflare's own team uses internally to move enterprise customers in hours instead of months, without a practitioner ever learning the full product suite first. That is the interface problem dissolving in real time, not a promise on a roadmap. That part is not where I would bet the shortage lands.

The shortage is what comes after. Every new agent-callable interface is a new attack surface, and it gets created faster than anyone can audit it. Proper auth, proper permission scoping, proper audit trails, none of that gets easier just because generating the interface got easier. If anything it gets harder, because now there are ten times more entry points into systems that used to have exactly one door, a human, in front of them. The scarce skill over the next three years is not the person who builds the agent-to-ServiceNow bridge. It is the person who can tell you, before it ships, every way that bridge can be walked through backwards.

The cybersecurity shortage is about to get worse, not better

This is the part I think most boards are underpricing, and it connects directly to the interface point above. Cybersecurity firms have been laying people off while telling investors it is AI efficiency. CrowdStrike $CRWD cut five percent of staff in 2025 citing AI-driven efficiency, then turned around and spent real money acquiring agentic security capability because the threat side of the equation moved faster than the defense side. That sequence should tell you everything. If AI made defense five percent more efficient, it made offense a lot more than five percent more efficient, and every new agent interface built this year is one more door for it to try.

Then a frontier lab built a model so good at finding zero-days that it was not allowed to ship publicly. Anthropic's Claude Mythos preview found working exploits in software that survived millions of fuzzing runs, at a cost of fifty dollars for a bug in OpenBSD and under a thousand for a full remote exploit chain in FreeBSD. That model stayed behind a defensive coalition, forty plus partners, a hundred million dollars in API credits, precisely because a lab that builds these models understood the offense to defense gap was already too wide.

Now play that forward two years, against an enterprise landscape full of newly built agent interfaces. A Chinese lab, or any lab outside that coalition, ships something in the same capability class in open weight form. The five person security team that used to be enough for a mid sized bank or an automotive supplier will not be enough anymore, and they will be defending far more surface area than they were two years earlier. Average breakout time for an attacker to move laterally inside a breached network is already measured in tens of minutes. When exploit discovery gets automated and cheap, that number does not slow down, it collapses.

Most companies I talk to have not internalized this. Banks, retail, automotive, traditional ecom, they are continuing on with headcount cuts and calling it AI transformation while barely touching their actual security posture. That is not a modernization story. That is a company betting its balance sheet on nobody finding the fifty dollar bug before they do. Palo Alto Networks $PANW and CrowdStrike $CRWD are the obvious plays if this thesis is right, and Cloudflare $NET fits the same bucket now that its own product is built to absorb the migration labor banks and retailers would otherwise need to hire for. Fortinet $FTNT and Zscaler $ZS sit downstream of the same demand for anyone screening the sector further. But the real shortage is human. Security engineers who understand both defensive tooling and how these models actually attack are going to be scarce and expensive, not cheap and automated.

The storage bill nobody budgets for

Underneath all of this sits something less dramatic but just as real. We learned this running our own stack. Every agent needs a place to remember things, and memory grows faster than anyone budgets for. Vector embeddings, session logs, context caches, tool call histories, it compounds. A knowledge base that fits on a laptop in month one needs a real storage tier by month six, and the query cost to search it well grows right alongside it. RAM matters for the working set, CPU matters for orchestration, but storage is the line item that quietly surprises you, and almost nobody plans for it before they hit it.

The market is already repricing around this. Pure Storage, now rebranded Everpure while keeping its ticker $PSTG, posted 13 percent revenue growth and 18 percent ARR growth this year and is explicitly positioning itself around AI-driven data management rather than plain disk capacity, a rebrand that only makes sense if the company believes agent memory is the next decade of demand. Nutanix $NTNX shipped an NVIDIA-certified storage stack built specifically to feed GPU clusters without bottlenecking them, scaling from 10 GB/s at 32 GPUs to 160 GB/s at 1,024 GPUs, and followed it in March with a full agentic AI stack that bakes vector databases and MCP server support directly into the storage layer. That is two public companies making multi-year bets that retrieval infrastructure, not raw compute, is where the next shortage of engineering talent and capital shows up.

The physical build nobody is arguing about

Away from the debate over whether agents replace jobs, there is a build happening that is not controversial at all, because it is physics, not economics. Every agent running at scale needs power, cooling, and interconnect, and none of those three are solved yet.

Rack density has gone from five kilowatts a few years ago to well over a hundred today, with next generation platforms targeting six hundred. Traditional air cooling cannot keep up, so liquid cooling has become the default line item for anything built for 2026 and beyond. Vertiv $VRT posted a project backlog above fifteen billion dollars and raised full year guidance on exactly this demand. Eaton $ETN is expanding from switchgear into liquid cooling and solid state transformers because the whole power delivery chain from the grid to the rack has to be rebuilt. Corning $GLW took a five hundred million dollar investment from Nvidia to expand fiber production tenfold, because moving data between GPUs is now as much of a bottleneck as the GPUs themselves. Arista Networks $ANET sits in the switching fabric that connects those clusters together.

None of this needs agents to be good at replacing marketers or engineers to pay off. It needs the industry to keep building agents at all, good or mediocre, and the physical bill comes due regardless of whether the software delivers on its promises. That is probably the safest two to three year bet in the entire AI trade, and it has almost nothing to do with whether prompt engineering ever gets solved.

What actually gets scarce

Put it together and the shortage list looks nothing like the popular narrative.

Security engineers who understand both the defensive tooling and the offensive model capability, because the five person team era is ending right as agent interfaces multiply the attack surface. Storage and retrieval engineers who understand how to keep a growing knowledge base fast and cheap to query, a cost nobody sees coming until month six. And on the physical side, the people who can build and commission power, cooling, and fiber capacity fast enough to keep pace with a buildout that is already outrunning the grid.

None of these are the jobs getting the headlines. None of them are "prompt engineer" or "AI whisperer." Building the interface will likely stop being the bottleneck. Securing it, storing what it remembers, and powering the machine underneath it will not.

I run a company that lives inside this exact tension every day. We are not ten times more efficient because we adopted agents. We built our own agent stack because we had to understand where the real cost and the real scarcity actually sit, not where the headlines say they sit. Interfaces are getting easier to build. Security teams are understaffed against a threat that just got a hundred times cheaper to execute. Storage costs more than anyone budgets for. And somebody still has to pour the concrete and pull the fiber.

The AI story that is coming is not fewer people. It is different people, doing work that has no shortcut yet, while the software everyone is excited about waits for the physical and human layers underneath it to catch up.

I do not know exactly which of these shortages bites first. My guess is security, because the cost of being wrong there is not a missed quarter, it is a breach. But I would not bet against storage and physical buildout labor being the quieter story that actually eats more of the budget over the next three years.

Research and tickers via stocks-llm. Explore the infrastructure buildout names: power, cooling and interconnect for AI data centers →. Or screen the cybersecurity side: security names against layoffs, AI efficiency and agentic M&A →.