There is a framework that has been quietly spreading through AI research circles that I find more clarifying than almost anything else written about where we are and where we're going. It borrows its structure from the autonomy levels developed for self-driving cars — L1 through L5 — and applies them to AI agents. The parallel is not cosmetic. The dynamics are remarkably similar: a gradual transfer of control from human to machine, with humans shifting roles at each level, until at the limit they are no longer operators but observers.
We are somewhere in the middle of this transition right now. And unlike the self-driving car analogy, where L5 remains largely theoretical, some domains of cognitive work are already fully there.
- The L1–L5 map — from basic responder to full organizational autonomy
- Why we are not at one level uniformly — arithmetic at L5, open-ended judgment still L2–L3
- The factory-floor parallel — artisan to operator to supervisor to observer
- Which categories of cognitive work have already been industrialized
- The physical leash — cognitive L5 with physical L1
- The observer problem — economics when execution is free
- What the manager role actually requires at L5
A Map of the Transition
L1 — The Basic Responder. The human is the operator. The AI executes specific, bounded tasks on command: summarize this text, draft this email, translate this paragraph. There is no initiative, no memory, no planning. The AI is a sophisticated tool, nothing more. Most people's first experience of AI — and still the dominant mental model — lives here.
L2 — The Collaborator. The AI connects to external tools: APIs, spreadsheets, calendars, databases. It can read, create, and modify documents. But it requires frequent human guidance to plan sequences of actions. The human is still deeply involved — less as a typist, more as a navigator who needs to give directions at every turn.
L3 — The Agent. The AI takes initiative. It plans and executes multi-step workflows over longer time horizons with limited intervention. The human becomes a consultant: setting the direction, available for questions, but not managing every step.
L4 — The Autonomous System. The agent operates with high autonomy, resolving its own blockers. A human is only needed to approve major decisions or when the agent hits something genuinely beyond its capacity. The human has become an approver.
L5 — The Organization. The agent manages full end-to-end projects independently. It sets sub-goals, coordinates resources, handles blockers, and delivers outcomes. The human sets the broad goal and then observes. No intervention required.
The levels aren't just a technical taxonomy. They describe a fundamental shift in the human role — from operator to collaborator to consultant to approver to observer. At each level, the human does less of the work and more of the directing.
Not One Level, But Many
The most important thing to understand about the L1-L5 framework is that we are not uniformly at any one level. Different domains of cognitive work are at radically different points on this scale — and have been for longer than most people realize.
Arithmetic, accounting, statistics: already L5 for decades. The industrialization of computation happened in the mid-twentieth century. Nobody manually calculates a standard deviation anymore. The machines handle it end-to-end. This is L5 — humans as observers — and it has been normalized so thoroughly that we no longer think of it as AI.
Structured code generation: approaching L4. Given a well-defined problem, modern AI systems can generate working code, run tests, identify failures, and iterate — with humans reviewing and approving rather than writing.
Open-ended reasoning and judgment: still L2-L3. For genuinely novel problems — strategic decisions, creative direction, ethical judgment — humans remain deeply involved.
The uncomfortable implication: the frontier is moving fast, and the categories that felt safely human six months ago are being reclassified.
How We've Been Here Before
The operator-to-manager transition is not new. It happened first in manufacturing, and the parallel is instructive.
Before the industrial revolution, a chair was made by a craftsman. One chair, one craftsman, labor from start to finish. The craftsman was the operator — doing every step of the work. Then came mechanization. The factory replaced artisanal production. The craftsman became the factory worker: still present, still necessary, but now operating machines rather than making things by hand. Then came further automation. The factory worker became the supervisor: overseeing the machines, catching errors, handling exceptions. Today, in the most advanced manufacturing facilities, humans barely touch the product. They monitor dashboards, approve interventions, manage the system. The car assembles itself. The human watches.
What happened to physical labor in the factory is now happening to cognitive labor in software. The progression is identical — artisanal to mechanized to supervised to observed. The only difference is the type of work being industrialized.
I've lived this transition in real time over the last few months. Two years ago, my workday involved writing code — directly, line by line, debugging and building. Today, I barely write code. I manage agents. I set objectives, review outputs, redirect when something goes wrong, approve before deployment. I have become, functionally, a manager of a software team where the team members are AI systems. The role shift that happened to factory workers over a generation happened to me in months.
Which Tasks Have Already Been Automated
Every wave of cognitive industrialization follows the same pattern: a category of mental work that previously required human attention is identified, modeled, and automated.
The first wave industrialized computation: arithmetic, record-keeping, statistical analysis. The second wave industrialized information retrieval and communication: search, databases, email. The third wave, happening now, is industrializing higher-order reasoning: the planning, sequencing, and structuring of complex tasks. The judgment about how to approach a problem. The synthesis of information from multiple sources into a coherent output.
What makes this wave different from the previous ones is not the sophistication of what's being automated. It's the proximity to what we thought of as the core of human intellectual work. Arithmetic felt mechanical even when humans did it. Reasoning feels like thinking.
The Last Leash
The L1-L5 framework describes cognitive autonomy. But there is a parallel stack for physical resources that rarely gets discussed — and it points at something important.
Every AI system today, however cognitively autonomous, is physically dependent. It needs servers, electricity, cooling, connectivity — all maintained by humans. There is a sense in which the most sophisticated L5 cognitive agent is still L1 on the physical resource axis: it cannot go source its own compute, cannot provision its own power, cannot maintain its own infrastructure.
This is the last leash. And it raises an interesting question: what would Physical L5 look like?
The answer, taken to its limit, involves something like nanotechnology: systems that can not only reason and plan but synthesize the physical materials they need on demand. The analogy that comes to mind is the 3D printer — a device that takes a digital specification and produces a physical object, layer by layer, from raw material. A food synthesizer would do the same for nutrition. A molecular assembler would do the same for arbitrary matter.
We are nowhere near this. But the conceptual direction is clear: the same trajectory that took cognitive work from L1 to L5 — progressively reducing human involvement in sourcing, maintaining, and operating the inputs — could in principle apply to physical resources too. The grid already does part of this elegantly: a machine draws however much electricity it needs, automatically, without human intervention. The machine fuels itself from the grid. Future systems may do the equivalent for compute, for materials, for everything.
The current AI systems are cognitively ambitious and physically dependent. As that physical dependency reduces, the nature of human oversight changes again. The question of what humans do when the machines handle both the thinking and the physical inputs is not yet answerable — but it is worth starting to ask.
What Happens When Everyone Watches
The L5 endpoint raises an economic question that the technical literature tends to skip: if humans become pure observers, what value do they add? And if they add no value, who pays them? And if nobody pays them, who buys the products the agents are building?
This is not a new concern. It has been raised at every wave of automation, and every previous wave resolved it the same way: automation eliminated certain jobs and created new categories of work that hadn't existed before. The bookkeeper displaced by the spreadsheet became the financial analyst. The travel agent displaced by the internet became the experience curator.
But there is an honest tension here. The previous waves of automation were slow enough that the economy could absorb them. This wave is faster. The gap between displacement and re-absorption may be longer and more painful than anything we've seen before.
Three possible resolutions, none of them certain:
- New categories of work emerge — as they always have, but stranger and harder to predict. What does it mean to be a professional director of AI systems? What judgment roles exist in a world where execution is free?
- Ownership of agents becomes the new capital — if your agents do the work, owning and directing agents is the economic activity. The question of who owns the agents, and how broadly that ownership is distributed, becomes the central economic question of the next generation.
- Redistribution becomes necessary — if productivity gains accrue to the owners of AI systems rather than to workers broadly, the political economy of that distribution becomes unavoidable.
The Role That Remains
In the factory analogy, the shift from operator to manager didn't make humans less important — it made them differently important. The manager sets the objectives that the machines optimize for. The manager catches the errors the machines can't see. The manager decides when the output is good enough and when it needs to be redone. The manager holds the judgment that the machine lacks.
The same is true for cognitive work. The L5 observer is not passive — they are doing something the agents cannot do: deciding what the agents should be trying to accomplish, evaluating whether they've accomplished it, and taking responsibility for the outcome.
The skills that matter at L5 are not execution skills. They are direction skills: clarity about what you want, ability to evaluate what you got, judgment about whether the gap matters, and wisdom about when to accept the output and when to push back.
These are the same skills that have always differentiated great managers from mediocre ones. The transition doesn't create a new type of human excellence — it reveals the one that was always most important, by removing the layers of execution that used to obscure it.
We are industrializing cognition. The artisanal phase is ending. What comes next, for the humans in the system, is the same thing that always comes next: a harder and more interesting role, requiring judgment that machines can't yet replicate, at a scale that was previously impossible.
That seems like the right direction to be moving in. The open question is whether we get there gracefully.