The Shrinking Layer

Hannah Arendt wrote in 1958 that the modern world faced "the prospect of a society of laborers without labor, that is, without the only activity left to them. Surely, nothing could be worse." She was not worried about unemployment. She was worried about a civilization that had organized itself entirely around work and then had work taken away, left with a freedom it had no idea how to use.

Every conversation about AI and jobs makes the same mistake Arendt was warning about. It treats the question as static. Can AI do this task today, yes or no? If yes, the human is displaced. If no, the human is safe. Draw the line, count the jobs on each side, publish the forecast.

The experiment we just ran broke that framing for me. Not because the results were dramatic. Because of what the results implied about time.

The self-evolution loop improved agent performance by 5.5% on tasks it had never seen, with zero human involvement. Business rules diverged 91% between two customers. The agent taught itself to parse informal text messages, enforce penny-level invoice precision, and select vendors based on reliability thresholds that no human programmed. All of this happened in what would be ten business days of production operation.

The question is not what AI can do. The question is what AI will learn to do next month that it cannot do today. And the month after that. And the month after that. The capability boundary is not a line. It is a wave, and it moves in one direction.

The human layer is defined by what the agent cannot yet do

This is the part the series has been building toward without saying plainly.

In The Retention Layer, I described how the self-evolution loop compounds customer-specific knowledge over time. In What an AI Agency Actually Needs, I described the human escalation layer: the team that handles the cases the agent cannot. I wrote that "the human layer makes the agent better, which makes the human layer smaller, which concentrates the remaining human effort on increasingly difficult edge cases."

I meant it as a feature. It is. But follow the logic all the way through.

If the human role is defined by the agent's current limitations, and the agent's limitations shrink autonomously every cycle, then the human role is not a stable category of work. It is a residual. It is whatever is left after the machine finishes learning this week.

That is not how jobs have ever worked. In every previous technological transition, the displaced tasks were visible and finite. The loom replaced the weaver. The spreadsheet replaced the bookkeeper. You could point at the boundary and it stayed where you pointed. The new boundary moves on its own.

Three layers, three timelines

I see three distinct phases of where human work goes as self-evolving agents mature. They are not sequential in the clean sense. They overlap, they vary by industry, and the boundaries are blurry. But the direction is consistent.

Layer one: humans define the game

Timeline: now through 2028.

This is where we are today. The agent executes. The human defines what good execution looks like. In ACP terms, the human writes the evaluation function.

Our experiment demonstrated this precisely. The agent improved by optimizing against a scoring rubric that measured vendor selection accuracy, margin calculation, invoice precision, and communication quality. Someone had to define that rubric. What counts as the correct vendor? What margin is acceptable? What tone matches this customer? Those are judgment calls grounded in operational experience that no amount of self-evolution can bootstrap from nothing.

This is the "the code is the easy part, the judgment is everything" argument from the agency post. It holds. For now.

The human value in this layer is domain expertise. The ex-operator who has dispatched ten thousand waste hauling jobs and knows what a good one looks like. The insurance adjuster who can tell the difference between a legitimate claim and a suspicious one without a checklist. The recruiter who knows why a candidate's resume does not capture what makes them right for the role.

These people are not writing code. They are not managing the agent. They are defining the criteria by which the agent judges itself. That is a new kind of work. It did not exist five years ago. It is the most important job in an agent-native company today.

The short-term forecast: companies that hire for domain expertise rather than technical skill will build better agents. The best evaluation functions will not come from engineers. They will come from the people who did the work before the agent existed.

This is not hypothetical. In April 2025, Shopify CEO Tobi Lutke issued an internal memo requiring employees to prove that AI cannot do a job before requesting new headcount. The burden of proof flipped. The default is now the machine. The human must justify their existence against it. That is layer one made corporate policy at a hundred-billion-dollar company.

Layer two: humans hold the line

Timeline: 2027 through 2031.

Even when agents can do everything operationally, someone must be accountable. Legally. Regulatorily. Ethically.

The sovereignty post described the audit infrastructure that makes this possible: version lineage, tenant isolation, evolution provenance. The Service Layer made the point that when the agent drives the decision, the vendor owns the outcome. These are not abstract principles. They are the foundation of a new category of human work.

The EU AI Act already requires human oversight for high-risk AI systems. Financial regulators require human sign-off on automated lending decisions. Healthcare requires a licensed professional to review AI-generated diagnoses. As agents move from back-office automation to customer-facing operations, the regulatory surface expands.

The human in this layer is not doing the work. They are not even defining the work. They are the person whose name is on the line when the work goes wrong. Their job is accountability, not execution.

This sounds like a demotion. It is not. It is a recognition that the hardest part of autonomous systems is not making them work. It is making them trustworthy. The human who understands the agent's evolution history, can explain its decisions to a regulator, and can intervene when the evaluation function drifts, that person is not a rubber stamp. They are the reason the system is allowed to operate at all.

The medium-term forecast: a new professional class emerges. Not AI engineers. Not domain experts. Agent operators. People whose skill is understanding what an autonomous system is doing well enough to take responsibility for it. Part auditor, part translator, part circuit breaker.

Layer three: humans decide what matters

Timeline: 2030 and beyond.

This is the layer I am least certain about and most interested in.

Today, the human defines the evaluation function and the agent optimizes against it. But what happens when agents become capable of proposing improvements to their own evaluation criteria? Not just "I should check vendor reliability before selecting the cheapest option" but "the evaluation function should weight customer retention more heavily than margin optimization because repeat customers generate higher lifetime value."

That is not execution. That is not even judgment about execution. That is judgment about what to value. And if the agent can do it well enough, the human role moves from "defining good" to something harder to name. Deciding what is worth doing. Choosing between competing values when they conflict. Determining which outcomes matter for reasons that cannot be reduced to a metric.

This is philosophy, not engineering. It is the question every major technological transition eventually forces: when the machine can do the work, what is the work for?

I do not have a forecast for this layer. I am not sure anyone does. But I am confident that the path to it runs through layers one and two, and that the timeline is shorter than most people expect. Our experiment showed 91% business rules divergence in two weeks. The capability frontier is not approaching slowly.

The rate of change is the story

Every previous displacement had a human-visible speed. Factories took decades to replace cottage industry. Spreadsheets took a decade to reshape accounting. Even the first wave of AI automation took years to meaningfully impact call centers and data entry.

Self-evolution changes the clock speed. The ACP loop runs daily. Each cycle, the agent attempts to get better. Many attempts are rolled back. But the ones that stick accumulate. The regression gate ensures the floor never drops. The learning curve only goes up, and it does not wait for a product manager to file a ticket.

Cognizant's 2026 workforce report quantified this mismatch. AI exposure across occupations is increasing at 9% per year, four and a half times faster than the 2% that was forecast just two years ago. The share of jobs with the lowest AI exposure shrank from 31% to 7%. Meanwhile, workforce retraining infrastructure in the US operates on multi-year funding cycles. The machine improves in sprint cycles. The humans reorganize in fiscal years.

In the waste hauling experiment, the agent went from being unable to parse "Hey Marcus need a 30 yard for a bathroom demo Thursday" to correctly extracting every field, selecting the optimal vendor, calculating margin to the penny, and formatting the quote in the customer's preferred communication style. Autonomously. In what would be two weeks of operation.

Two weeks is not a labor market disruption timeline. It is a sprint cycle. The agent is improving at the speed of software deployment while the humans around it are adapting at the speed of career transitions, organizational restructuring, and regulatory evolution.

That mismatch, the gap between how fast the agent learns and how fast humans reorganize around what it has learned, is the actual disruption. Not the capability. The velocity.

What I am building toward

I am writing this because the ACP series has, until now, treated the shrinking human layer as a clean optimization story. Smaller teams, higher margins, better outcomes. All true. But incomplete.

The founders building agent-native companies need to think about this now, not after the first wave of displacement makes it urgent. Not just what the agent can do, but how fast it will learn to do more. Not just how many people you need today, but what those people will do in eighteen months when the agent has had thirty-six more evolution cycles to absorb the work they currently perform.

The next post in this series will go deeper into the question that layer three raises. When autonomous systems can improve their own evaluation criteria, where does human purpose live? That is not a workforce planning question. It is a civilizational one.

But it starts here, with the recognition that the timeline is measured in sprint cycles, not decades. And the clock is already running.

In 1965, the mathematician I.J. Good, who had worked alongside Turing at Bletchley Park, wrote: "The first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control." The proviso is the part worth sitting with.