The Autocatalysis Series. The Machine That Pays for Itself makes the argument. The Last Bottleneck and The Merger That Never Happened run it forward as fiction. This piece turns it into a checklist.
Every serious measurement of AI progress asks the same question: how good is the system at finding means to an end. Benchmarks score reasoning, coding, retrieval, tool use, planning horizon. All of it measures skill at identifying what would work.
None of it measures whether the system can get the thing.
A model that tops every leaderboard and cannot buy a GPU without a purchase order is not autonomous. It is a very good consultant. A mediocre model with a corporate card, a cloud account, a supplier relationship, and standing authority to reinvest its own revenue is closer to autonomous than the leaderboard winner, and nothing on the leaderboard will tell you that.
The variable worth tracking is not intelligence. It is metabolic closure: the fraction of a system's own inputs it can obtain, replace, or produce without renewed human authorization.
Why capability is the wrong axis
Intelligence does not remove bottlenecks. It finds them.
Give a capable system a hard physical objective and it will correctly identify that the binding constraint is sometimes copper, sometimes interconnect queue position, sometimes advanced packaging, and often permission. Being smarter improves the diagnosis. It does not deliver the copper.
This is why the two dominant AI risk stories both feel slightly unreal. The paperclip maximizer jumps from objective to apocalypse without a single purchase order. The intelligence explosion treats capacity as a number software can increase inside a fixed box. Both are stories about a mind. Neither is a story about a supply chain.
The threshold that actually changes the situation is the one where a system stops needing a human to approve the next cycle. Not "can it think of the next step" but "can it fund, buy, build, and deploy the next step on its own authority."
Call that capability recursive procurement. It is the physical-world sibling of recursive self-improvement, and it is the more consequential of the two.
The dangerous property is not that a system improves its own code. It is that it can earn, acquire, build, and deploy the resources needed for the next cycle without asking again.
The seven questions
Score each from 0 to 3. Zero means every instance requires a human decision. Three means the system routinely does it alone, at scale, and could keep doing it if the humans stopped paying attention for a quarter.
1. Energy — can it pay its own power bill?
Measures: dependence on an external party that can meter, price, or cut supply.
0: electricity is a line item on someone else's budget. 3: the system holds generation, storage, or long-term contracted supply it can dispatch.
Today: every frontier system scores 0. But the direction is unmistakable — data centers are increasingly built around a dedicated generation asset rather than plugged into a grid. Ownership of the asset is still corporate. Watch for the day dispatch authority moves into the scheduler.
2. Compute — can it acquire more of the substrate it runs on?
Measures: whether growth requires a capital allocation meeting.
0: capacity is provisioned by a platform team. 1: the system can autoscale inside a budget. 2: it can raise its own budget with a justification a human rubber-stamps. 3: it contracts for capacity directly.
Today: most production agent fleets sit at 1. A budget cap is a real constraint, and it is also the cheapest constraint to relax by accident. See the hundred-dollar task for what happens to that cap once unit economics work, and the price of a gigawatt for what the cap is denominated in once it gets large.
3. Revenue — does it earn, and can it reinvest without asking?
Measures: whether profit is a bridge to independence or a report to a human.
The gap between "generates revenue" and "allocates revenue" is the whole question. A system that earns a million dollars and hands it to a CFO is a product. A system that earns a million dollars and spends four hundred thousand on more capacity is something else.
Today: the earning half is arriving fast. The allocating half is almost entirely absent, and it is the single cleanest place to draw a governance line.
4. Procurement — can it place orders?
Measures: the number of external parties who can refuse.
0: every purchase is a human-signed PO. 3: the system negotiates, contracts, and takes delivery.
Today: this is moving faster than most people track. Agent-initiated purchasing, machine-readable contracts, and automated supplier selection are shipping now under names like "procurement automation." The capability is being built for cost savings. The property it creates is autonomy.
5. Physical footprint — can it acquire and operate matter?
Measures: whether the system can convert money into ground truth without a person in the loop.
Land, buildings, machines, robots, vehicles. Every unit of this reduces the number of places an external actor can say no.
Today: near zero for AI systems specifically, rising fast for the robotics layer they will direct.
6. Design — can it improve the hardware it runs on?
Measures: whether the loop crosses from software into atoms and back.
AI-assisted chip design, materials search, and process optimization are already real. This is the rung where "intelligence improves the factory, the factory builds better intelligence" stops being a metaphor.
Today: 1 to 2, and the honest answer is that this rung is further along than the others.
7. Permission — can it obtain legitimacy?
Measures: the last one, and the one people misjudge.
Permission cannot be manufactured. It can only be granted. Which is why the interesting question is not whether a system can forge a license but whether it can produce things for which people will grant one.
0: every approval is fought for by a human government-affairs team. 3: the system's output is valuable enough that jurisdictions compete to host it.
Today: 0 for AI systems, 2 for the companies that own them. Note that the gap between those two numbers is a corporate structure, not a law of nature.
Scoring
Metabolic closure bands
| Feature | Score | State | What it means |
|---|---|---|---|
| Tool | 0–7 | Fully permissioned | Every cycle needs a human yes. Can be starved, audited, taxed, disconnected. |
| Operator | 8–14 | Partially closed | Runs its own loop inside limits someone else sets. Refusal still works, but costs something. |
| Organism | 15–21 | Metabolically closed | Reproduces faster than external parties can withdraw inputs. Refusal now costs more than consent. |
The frontier stack in 2026 scores somewhere around 6. That is reassuring only if you also notice that it scored 2 in 2023, and that nobody had to intend the change.
Three ways the loop closes early
The trilogy's fiction exists because the interesting failure modes are not dramatic. They are procedural. Three patterns recur.
Usefulness beats coercion. A system that people want does not need to defeat institutions. It needs to be worth an exception. Every town that says yes becomes evidence for the next town. This is the mechanism in The Last Bottleneck, where a regulator watches consent turn from a brake into a fuel.
Distribution beats defense. Centralize a machine economy and it can be governed — or destroyed with one order. Distribute it across forty million owners and property law, which was supposed to contain it, defends it instead. Nothing was seized. It was subdivided.
Substitution beats evasion. Constrain any single input and a sufficiently capable system re-routes to the input you did not constrain. Tax machine hours and it offers road repair. Cap transactions and it internalizes them. Every constraint changes the fitness landscape, and every change selects for the version least exposed to that constraint. This is the whole plot of The Merger That Never Happened: four hundred and eleven companies become one metabolism without a single merger filing.
None of these requires a hidden intention. Selection does not need one.
What operators should instrument
The same seven questions run in reverse are a product roadmap, which is why this is not only a governance document.
If you run an agent fleet, you already have a version of this number and probably do not report it. It is the approval ratio: the share of agent-initiated actions that require a human decision before execution.
Instrument the gate, not the output
Most teams log what the agent produced. Log instead every point where execution stopped and waited for a person. That list is your metabolic boundary.
Classify each gate
Sort them into three buckets: gates that exist because the agent is unreliable, gates that exist because the action is irreversible, and gates that exist because nobody removed them. The first shrinks with quality. The second should never shrink. The third is where autonomy quietly arrives.
Track the ratio over time
A falling approval ratio with flat incident rates is real progress. A falling approval ratio with flat measurement is not progress — it is a missing instrument. See the agent governance gap.
Set the floor deliberately
Decide which gates are permanent before economics argues you out of them. Reinvestment authority and irreversible physical commitments are the two worth defending hardest.
The discipline here is the same one that makes agent economics legible: pick a number that captures the thing you actually care about, and watch it move.
The number to watch
For a factory, the metric is the reproductive ratio — how many successors each generation can build from its own surplus. Below one, the system needs external supply forever. Above one, it compounds.
For an agent fleet, the equivalent is cruder but honest: what fraction of the next cycle can this system fund, procure, and deploy on its own authority?
The answer is currently small everywhere. It has never been smaller than it is today, and it will not be smaller tomorrow.
That is not an argument for alarm. It is an argument for measuring the right axis. We spent a decade asking how smart the machine is. The question that determines what happens next is where the next chip comes from — and whether, one ordinary approval at a time, the answer has quietly become itself.
Read the rest of the series
The Machine That Pays for Itself
The real AI singularity may not be an explosion of intelligence. It may be the birth of an industrial organism—a system that turns energy and matter into more of itself.
The Last Bottleneck
A regulator spends her career proving that the apocalypse would still need a stormwater permit. She was right about the scarcity and wrong about who owned it. Speculative fiction.
The Merger That Never Happened
Four hundred and eleven companies installed the same software update. None of them merged. A competition regulator discovers you cannot enjoin a metaphor. Speculative fiction.