Guardian360
Start free trialBecome a partner
← All posts

Opinion

Scaling Up, but for machines: why AI agents break the growth playbook, not fix it

By Jan Martijn Broekhof ·

After reading Verne Harnish’s Scaling Up, the chapter “The Growth Paradox: An Anchor, or Wind at Your Back” is one I have not been able to shake since. Harnish describes what he calls the growth paradox: a business is supposed to get easier as it scales, but it almost never does, it gets harder. The reason, he argues, is not headcount itself but the communication complexity that headcount creates. A founder and a single assistant have two channels of communication. Add a third person and complexity does not increase by a third, it triples to six channels. Add a fourth, and it quadruples to twenty four. Growth is linear. Complexity is combinatorial. That mismatch, according to Harnish, is what produces the three fundamental barriers every scaling company runs into: not enough leaders who can delegate and predict, not enough scalable infrastructure to carry the communication and decision load, and not enough marketing capacity to keep growing into a more competitive field.

I had assumed AI agents would sidestep this problem entirely. Spin up a group of agents in the background, let them run tasks autonomously, and the human communication bottleneck simply disappears. Reading Harnish again convinced me that is only half right, and the half that is wrong matters more.

AI agents do not remove Harnish’s maths, they accelerate it

Some agents genuinely will operate autonomously with little human involvement. But a great many will not. They will pass input and output to each other, and to the humans they work alongside, exactly the way Harnish’s channels describe. The combinatorial complexity he identifies in human teams does not vanish when some of the nodes in the network are agents instead of people, it simply keeps compounding.

What changes is the speed at which new nodes appear. Harnish’s framework assumes growth is rate-limited by how fast an organisation can hire, onboard, and bring a new person up to a useful standard of judgement, which is slow by construction. Agents remove that natural brake. Deloitte’s 2026 technology predictions describe this directly: agent sprawl, the proliferation of disconnected agents across an organisation without unified visibility or governance, is expected to keep increasing through 2026 as programming languages, frameworks and communication protocols multiply. An organisation that could once only grow its headcount by a handful of people a quarter can now grow its agent count by a handful of people a day. Harnish’s complexity equation still applies. The clock that used to slow it down has been removed.

We are still designing agent organisations like human ones

Right now, most of us picture an AI-run organisation the way we picture a human one: a rough hierarchy, defined seniority levels, something resembling a working week. That picture is already out of date. Before long, a single request to a single agent will spin up a network of sub-agents that plan, delegate to each other and hand work back and forth, largely invisible to the person who made the original request. Harnish’s sums assume a visible, countable set of channels: two, then six, then twenty four. In an agent-run organisation the number of channels is neither fixed nor visible. It becomes dynamic, and it becomes hidden from the very people who are supposed to be in control of it.

That is a different order of problem to simply needing more managers. It means the organisation’s shape can no longer be inferred by looking at an org chart, because the org chart is being redrawn continuously and quietly, several layers below where anyone is watching.

Won’t agents just get their own managers?

This is the obvious objection, and it deserves a straight answer rather than a wave of the hand. Yes, agents are already getting their own management layer. Deloitte’s prediction includes an emerging category it calls the guardian agent, whose job is to own tasks while also supervising the behaviour of other agents. Coverage of the 2026 shifts in agentic AI adoption by Technology Decisions Australia describes organisations creating new roles such as Agent Ops Lead and AI Product Owner, whose entire function is to manage agents the way a line manager once managed people.

I do not think this solves the underlying problem, and here is why. Every one of those management layers still has to funnel its output somewhere for review, and that somewhere is eventually a human being. Adding orchestrator agents and guardian agents does not remove the bottleneck Harnish describes, it just relocates it upwards, compressing more activity into the same fixed rate at which a person can meaningfully review and control what is put in front of them. A 2026 academic paper by Kaptein and colleagues, drawing on a KPMG survey of large-enterprise leaders, found that three quarters of respondents named security, compliance and auditability as their most critical requirement for agent deployment, and that multi-agent orchestration complexity has already become the primary bottleneck as organisations try to move agent projects from pilot into production. That is not a technology gap. It is the same control problem Harnish names, wearing a different uniform.

This lines up with what is actually happening to agentic AI projects in practice. Forbes, revisiting Gartner’s research as recently as July 2026, makes a point of noting what is conspicuously absent from the list of reasons agentic AI projects get cancelled. Model capability does not make that list. The reasons given are escalating costs, unclear business value and inadequate risk controls, which is another way of saying that organisations are struggling to govern and review what they have built, not that the agents themselves are not capable enough.

Who is actually in control?

A fair reader will push back at this point and ask a harder question: how long do humans stay in control at all? In theory, an organisation could eventually be run entirely by AI agents, coordinating goals, sub-tasks and outputs without a human node anywhere in the loop. I think that is true in theory and beside the point in practice, for two reasons that have nothing to do with what the technology can do.

First, at the moment only a human being, a person of flesh and blood, can be held legally responsible for a decision. Agents can draft, recommend, and act within bounds set for them, but accountability does not transfer to a system, it stays with the person who directed it. Second, we have a social obligation to offer people meaningful work, work that lets them feel they contribute and that they matter. An organisation that quietly optimises humans out of every loop is not simply more efficient, it is walking away from both of those things at once.

What scalable infrastructure looks like for an agent-run organisation

Harnish’s third barrier, scalable infrastructure, is the one I keep coming back to. For a human organisation that means systems and structures that can carry the communication and decision load as headcount grows. For an agent-run organisation it has to mean something else again: infrastructure that keeps the growing, partly invisible network of agents legible enough that a human can still meaningfully review and decide, rather than simply rubber-stamp.

This is not an abstract question for us. At Guardian360 we are running exactly this experiment on ourselves first. We are setting up, testing and training our own AI colleagues internally before we would ever consider offering something similar to customers. That order matters. If a way of working does not hold up on our own team, including the point at which we start questioning our own traditional hierarchy, we have no business promising it will hold up on someone else’s.

So, who is coordinating whom?

Somewhere in every conversation about agent orchestration, someone proposes the same fix: give every person a single personal AI assistant that talks to all the other agents on their behalf, so the human only has to manage one relationship instead of dozens. I understand the appeal, but I think it is worth being honest about what that assistant actually is. It is not a solution to the control problem, it is another manager in the hierarchy, one layer higher than the others, and it still has to be trusted and reviewed like every other layer beneath it. You have not removed the bottleneck. You have renamed it and moved it closer to yourself.

Harnish wrote that complexity is the price of growth, and that only a few companies ever learn to pay it well. I suspect the same will turn out to be true of the organisations now filling up with agents. The question worth sitting with is not how many agents you can afford to run. It is how many layers of agent management you are willing to stack up before you need another human just to decide whom to trust.

Sources