Agency ⊥ intelligence · a scored, testable ladder

The Agency Ladder

Not "is it intelligent?" — that is a property the evaluator assigns, and the field cannot benchmark what is assigned rather than possessed. The system-intrinsic question is: does it spend its own energy to hold its own boundary, on its own power? Here is a five-rung ladder for that — and where today's systems actually sit. Most score zero.

Charlot Lab, Institute for Physical AI @ BMI

Draft · 2026-07 · companion to One Energy, Both Roles · every placement is argued from a measured result or an explicit open problem

Ashby ultrastability · Di Paolo adaptivity · autopoiesis Scored by J/VT — joules per viability-second Honest · in simulation · open problems named

Intelligence is an attributional predicate — it lives in the evaluator's task frame, which is why no one can build a benchmark for it. Agency is different: it is a property of the system, and it has a sharp discriminator — remove the human; does it persist? The boundary that fails when the power stops belongs to whoever owns the power, so a model in a datacenter has borrowed its viability. We make the discriminator a five-rung ladder (essential variables → adaptivity → second-order reorganization → self-production) and score real systems against it. A frontier LLM with full tool access scores 0; a Roomba scores 1. An on-device, certificate-gated controller — one that spends its own microjoules to hold its own energy boundary, and senses the tendency toward that boundary and acts before it is crossed — reaches rung 2, and one step into 3, and unlike the rest, its placement is provable, on a metric the field lacks: joules per viability-second held under disturbance.

1 · The wrong question, and the right one

The field is stuck measuring intelligence, and it is stuck for a structural reason: intelligence is assigned by an evaluator inside a task frame, not possessed by the system. You cannot benchmark a property that is attributed rather than intrinsic — chase it and the epistemics stay broken, which is exactly what has happened.

Agency admits a system-intrinsic test. The discriminator that does the most work in one line: if removing the human changes nothing about whether the system persists, there is no agency there. Almost everything currently labelled "agentic" fails this immediately — an LLM in a while-loop with tool calls has no essential variable of its own, no energy store of its own, an entirely exogenous setpoint, and it does not persist between invocations except in a buffer someone else rents. That is delegation, not agency. And the boundary that counts is the one that fails when the power stops: for anything running in a datacenter, that boundary belongs to the datacenter. Which is why, for a physical agent, running on the robot's own power, on-device, is not an efficiency nicety — it is a necessary condition for the boundary to be the robot's own.

2 · The ladder

Five rungs, drawn from the sharpest formalizations anyone has produced and mostly forgotten — Ashby's ultrastability1, Di Paolo's adaptivity2, the operational-closure half of autopoiesis3:

Rung	Test	Nothing above it is free
0	exogenous setpoint, exogenous power	delegation — a tool driven on someone else's goal and budget
1	an essential variable it can lose, on its own store	it can die, and it acts to not — the first real stake
2	senses the tendency toward the boundary; acts before it is crossed	a cliff becomes a gradient — Di Paolo's adaptivity, the difference between a system that dies and one that behaves
3	second-order reorganization — replaces its controller when the current one fails	Ashby ultrastability; every RL system tunes parameters inside a fixed architecture — that is the first loop only
4	self-production — makes the components that make it	operational closure; a program does not produce its own substrate

A frontier LLM with full tool access — vastly more capable than a Roomba. Which sits higher on the agency ladder?

owns an essential variable it can lose senses the tendency, acts before the crossing replaces its controller when one fails produces the components that produce it

Interactive figure. The ladder is gated, not additive: nothing above a rung is free, so a system sits at the highest rung whose test it passes and every test below it. Tick capabilities and watch where a system lands — skip a lower rung and the ladder refuses to advance. This is why raw capability moves nothing: a frontier model with full tool access still runs on an exogenous setpoint and someone else's power, so it never leaves rung 0, while a Roomba that seeks its own charger is already at 1.

Show the computation

rung = highest r such that tests 1..r ALL pass   (gated, not a score)
1 essential variable it can lose, on its own store      → a real stake
2 senses the tendency, acts before the boundary         → adaptivity (Di Paolo)
3 second-order reorganization: replaces the controller  → ultrastability (Ashby)
4 self-production of its own components                 → operational closure
skip a lower rung and the ladder does not advance — the gate is the point

Scored against real systems, the table is correct and insulting to precisely the right people:

System	Rung	Why
Frontier LLM + full tool access	0	exogenous setpoint + power; nothing of its own persists; dead the instant the datacenter is
Industrial robot arm	0	someone else's setpoint on someone else's power budget
Thermostat	0	the setpoint is exogenous and it does not pay to hold it
Roomba	1	it seeks its own charger — an essential variable (charge) it can lose
EFA certificate-gated controller, on-device	2 → 3	its own energy boundary; senses the tendency and acts before the crossing; the fallback switch is one step of ultrastability — argued below, and provable
Any living cell	4	it produces the components that produce it — the criterion nothing synthetic meets

3 · Where the certificate lands — and why it is provable, not asserted

An on-device controller gated by an energy certificate4 holds an essential variable it can lose — the energy $V$ must stay in the safe set $S=\{V\le c\}$ or the closed loop diverges (rung 1). It does not wait for the crossing: the discrete barrier admits an action only if its actual next state stays in $S$, so it acts on the tendency, before the boundary — a cliff turned into a gradient (rung 2). And it switches controllers under threat: when the task policy would leave $S$, the certificate vetoes it for an energy-shaping fallback — a first, real step of second-order reorganization (reaching rung 3). It runs in microjoules on the robot's own power, which is what makes the boundary its own at all.

Unlike every other row, this placement is not a claim — it is measured, on a metric the field does not have: J/VT, joules per viability-second held under disturbance. Under an adversary trying to drive the body out of $S$, a puppet that commits the policy holds viability 26% of the time at 347 W; the certificate-gated agent holds 100% at 57 W — cheaper and safer, not a trade. On a real robot arm (SO-101) the agent runs at 0.25 W with zero crashes against a single-vector policy's 7.47 W and 36 crashes. And it earns rung 2 in the hard case: given only intermittent, noisy, position-only sensing, an observability-aware version sizes its margin to a Kalman estimator's own uncertainty — when an occlusion hits and it can no longer see, that uncertainty grows on its own, the margin widens, and the body slows down rather than trusting a state it cannot estimate. Safe because it knows it cannot see.

So the row reads: rung 2, one step into 3. Not rung 3 in full — the reorganization is a switch to a fixed fallback, not a learned new controller. Not rung 4 — nothing synthetic produces its own substrate. But it is strictly higher than anything the industry ships, and it is the only row whose score comes with a proof.

What is genuinely unsolved — stated, not hidden. Boundary-drawing is not solved: a Markov blanket can be drawn in many defensible places around the same system, and the agency verdict changes with the cut — every criterion resting on "its own boundary," including this one, inherits that. Nesting is not solved: agency appears at mitochondrion, cell, tissue, organism, and colony at once, and nothing tells you which level is the bearer — the ant colony is the strongest agent in any such list and no individual ant is. And the obvious objection — agency without a for-what is kudzu; cancer has excellent agency — is real; our answer is not to smuggle intelligence back in but to keep it as the orthogonal axis (agency ⊥ intelligence): the generative vector supplies the for-what, the certificate supplies viability, and neither is the other. Finally, this is all in simulation; the deployed peer that runs this philosophy on real hardware across continents (physics + uncertainty baked in, on-device, robot-agnostic) is ahead of us on reality — our contribution is the formal object and the scoreboard, not the fleet.

References & lineage

W. Ross Ashby, Design for a Brain, 1952 — ultrastability: essential variables plus a second-order loop that reorganizes the controller when they leave bounds. Requisite variety gives the hard lower bound on required sensing.
Ezequiel Di Paolo, Autopoiesis, adaptivity, teleology, agency, 2005 — adaptivity as the necessary repair to bare closure: sensing the tendency toward the boundary and acting before it is crossed.
Maturana & Varela, autopoiesis (the operational-closure half only); Still, Sivak, Bell & Crooks, Thermodynamics of Prediction, PRL 2012 — the dissipation floor under a modeling agent.
Companion: One Energy, Both Roles and Energy Is the Certificate, Charlot Lab — the certificate and the J/VT scoreboard the rung-2 placement is argued from.

The reproducible probe scripts behind every number here (the certificate gate, the J/VT scoreboard, the observability-aware estimator loop) are the Institute's on-device research record. Agency is measured, not attributed — which is more than "intelligence" has ever managed.