Source
Barandiaran, Di Paolo & Rohde (2009), LIT-566, read in NOTE-458: §2 (the three conditions), Table 1 (the cases), §4 (the generative definition), footnote 3 (emergent regulation), §6 (the consequences for models).
What was actually shown
The paper argues from cases. It does not measure anything.
Three conditions. Individuality: the system distinguishes itself from its environment, and an observer does not do it for the system. Otherwise the explanation regresses through the observer. The paper quotes Jonas: an identity “owned by, not loaned to its subject” (§2.1). Interactional asymmetry: the system is the source of modulations of its coupling with the environment. This is defined as the system changing a subset of the constraints on an otherwise symmetrical coupling (Eqs. 1–3), not as being the bigger energy source or the stronger statistical influence. Both of those fail on clear cases, the gliding bird and the diver against the faller (§2.2). Normativity: the coupling is regulated against a condition that can fail and that the system’s own organisation sets (§2.3).
Individuality is necessary and not sufficient. Table 1 sorts cases.
- A cell undergoing passive osmosis is an individual, “producing and maintaining its organization including its membrane”. The process may even serve it. But the process is physically unconstrained and not caused asymmetrically by the cell, so it is not action.
- A kitten warmed by its mother is an individual, and the coupling meets the kitten’s norm of staying within viable temperature. But the mother drives the coupling, so the kitten is not acting.
- A person with a Parkinsonian tremor is an individual and the energetic source of the movement. The movement answers to no internally generated norm, so it is not action.
- A bacterium in metabolism-dependent chemotaxis meets all three, and it acts.
The authors claim sufficiency more weakly: they “find it very difficult to conceive of any system that jointly fulfills these three criteria and it is not an agent”.
The generative definition. A system S is an agent for a coupling C with an environment E iff S is an open autonomous system and S modulates C adaptively, that is, so as to maintain some of its own processes (§4). An autonomous system is a network in which every process depends on at least one other and enables at least one other, and which would run down in isolation. The three conditions follow from this. Autonomy here is Varela’s, explicitly “not restrict[ed] or reduce[d]” to metabolism, and its processes may be neural, sensorimotor or social. Footnote 3 adds that the modulation must be emergent. “Subsystems that regulate according to a fixed norm independent of the agent’s current state … act as external constraints on the system.”
What this does not say
- It does not say that an agent is a self or a person. The paper’s “self” is the autos of an autonomous organisation, and its “autonomy” is organisational. ADR-024 keeps agent, individual, self and person apart. This account separates the first two and says nothing about the other two. A bacterium is an agent on this account, and no one in the record claims it is a subject or a person.
- It is not about self-governance. Personal autonomy, governing oneself by standards that are one’s own, is the subject of THEORY-029, THEORY-054 and THEORY-047. The same word names a different thing. Nothing here bears on whether a motive is one’s own.
- It is not about any of the four unities. It does not say how many control systems act as one (behavioral-integration). Footnote 3’s exclusion of a central controller has the same shape as THEORY-056’s denial of a distinct regulating system, but neither paper cites the other, and the resemblance is the record’s.
- The asymmetry criterion is incomplete by its authors’ account. “What counts as a parametric modulation and what as a coupling between variables is often a matter of choice for the modeler” (NOTE-458, Limitations). That is the same kind of modeller-dependence that THEORY-073 finds in Markov blankets, and the authors name it as the open problem.
- Life is sufficient, not necessary. The definition is meant to admit non-living agents. It does not say what any present machine is. Consequence (b) says that systems “that only satisfy constraints or norms imposed from outside (e.g. optimization according to an externally fixed function) should not be treated as models of agency”. That is a remark about modelling. It is not a finding about any system, and it is no instruction for machine-learning practice.
Connections
- THEORY-072 (organisms versus dissipative structures) is not extended. Both ground norms in a self-maintaining organisation, and this paper cites Mossio et al. 2009 (LIT-570) for that. But its autonomy is a network of mutually enabling processes, not closure of constraints, and it makes no use of differentiation. Its line against a flame is a different one. A flame may well be self-maintaining in the broad sense, but it does not adaptively modulate its coupling. So this account neither needs nor refines the demarcation in THEORY-072. The two are independent conditions, and an organism in that account’s sense is not thereby an agent in this one.
- THEORY-073 (blankets do not individuate). Both conditions this paper puts first, individuality the system defines for itself and asymmetry that is not a statistical measure, are what a Friston blanket lacks on that account. The blanket is drawn by the modeller, and its arrow structure is stipulated in the graph. Bruineberg et al. borrow this paper’s term “interactional asymmetry” for that arrow structure (LIT-603). On this account, the arrow structure is not interactional asymmetry.
- THEORY-043 (minds of agents). That account reports that its readings agree that “agency survives both directions”: Block’s homunculi-head, List’s group agents (LIT-401) and Levin’s nested collectives. The agency they agree on is the thin functionalist or cybernetic kind. On this account, the agency is not cheap. Minsky’s mindless agents and Brooks’s behaviours have norms that their designers set, and no individuality of their own. Whether a firm or a court meets the three conditions is a substantive question, which neither List nor this paper answers. THEORY-043’s verdict on consciousness is untouched. Its premise that agency is not in dispute holds for the thinner notion only.
- Levin (LIT-439, NOTE-341). Levin’s Self is the set of parts that pursue a goal together, where a goal is the set point of “a feedback system that operates to maximize some specific state of affairs”. NOTE-341 records that this makes any homeostat a Self, and that the paper accepts it. On this account a homeostat is not, by that alone, an agent. Its norm is fixed from outside, which footnote 3 counts as an external constraint, and it does not make itself an individual. The two criteria also run in opposite orders: Levin individuates by the goal, and this account needs the individual first. ADR-024 already reads Levin’s Selves as agents and individuals. This account says that on the organisational reading, many of them would be neither. Levin’s account is stated in THEORY-583, a rival: the two cannot both be right about the homeostat.
- Participatory sense-making (LIT-571). The same group extends “sense-making”, a deviation from a self-generated norm becoming significant, to social interaction. It requires that the participants remain autonomous agents in this sense.
- Biological Autonomy (LIT-536) has a chapter on agency (chapter 4). It is unread, and the paper’s autonomy is wider than that book’s metabolic one.
Note of 2026-10-29: what holds it back
Kept Proposed by the tenth status review. NOTE-458 read the paper in full, and it argues from cases. The test its promote_when names is not in the record: the three conditions, with an asymmetry criterion that does not depend on how the modeller splits parameters from variables, applied to explicit models (a protocell, an evolved controller, a homeostat) and matching Table 1 without being tuned to it. The authors call that criterion incomplete. The disagreement with THEORY-583 over the homeostat is unresolved. What the paper proposes is now THEORY-1119, Active, and the claims that need only the proposal ground on it.