Contribution
Before this paper, “agent” in adaptive behaviour and robotics was defined by lists (perceives and acts; pursues goals; acts on its own behalf) whose terms presupposed what was to be defined. The paper replaces the lists with three conditions and then a generative definition: a type of organization from which the three conditions follow, stated without using “agent”, “action” or “goal”.
Key insight
Being an agent is not a relation between a pre-given system and its world. The system has to make itself a system (individuality), and the norms its actions answer to have to come from the same organization that makes it. Asymmetry then follows: the organization modulates its coupling because its own continuation depends on it.
Assumptions
- Intuitive adequacy is a constraint. The conditions are tested against ordinary uses of “action” (Table 1: gas, osmosis, tremor, warmed kitten, chemotactic bacterium).
- Coupling is symmetric at base. dS/dt = F_Q(S,E), dE/dt = G_Q(S,E) (Eqs. 1–2); asymmetry is modulation of a subset p of the constraints Q over an interval T, Δp = H_T(S) (Eq. 3). The constraints may be non-holonomic, so they cannot be re-described as variables.
- Autonomy in Varela’s sense, applicable beyond metabolism: component processes may be chemical, physiological, neurodynamic, sensorimotor or social habits.
- Emergent regulation (footnote 3). Modulation must not be attributable to a subsystem acting as a central controller with a fixed norm, which would be an external constraint on the system.
Key results
- Individuality (§2.1). An agent must distinguish itself from its environment without an observer, or explanation regresses through observers (Jonas 1968: identity “owned by, not loaned to its subject”).
- Asymmetry (§2.2). Energetic criteria fail for the gliding bird; statistical-correlational criteria (citing Lungarella et al. 2007, Seth 2007) fail for diver versus faller. Asymmetry is defined as the capacity to modulate the coupling at some times, at the extreme “just capable of halting a coupling”.
- Normativity (§2.3). Norms “cannot be deduced from universal laws alone”; failure must be possible. Tremors are not actions.
- Joint necessity and sufficiency (§2.4). Individuality is a precondition of the other two. The authors “find it very difficult to conceive of any system that jointly fulfills these three criteria and it is not an agent”.
- Definition (§4). S is an agent for coupling C with E iff (1) S is an open autonomous system: a network in which every process depends on at least one other and enables at least one other, which would run down in isolation; E is the set of processes that affect and are affected by S; S depends on conditions in E; and (2) S modulates C adaptively, i.e. so that some of its constituent processes are maintained.
- Life is sufficient, not necessary (§3–4). Protocellular metabolism plus membrane plus metabolism-dependent chemotaxis satisfies the definition.
- Spatio-temporality (§5). Agency is temporally structured (actions have onset, acceleration, consummation and cadence, after Di Paolo 2005) and spatially situated; a situated reactive controller can solve tasks a non-situated one cannot. The agent’s perspective, not the observer’s, fixes its spatial and temporal world. This section is declared “less rigorous”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Individuality, interactional asymmetry and normativity are jointly necessary and sufficient for minimal agency | moderate | conceptual analysis against cases, Table 1 |
| C2 | Neither energetic nor statistical-causal measures suffice to define asymmetry | moderate | counterexamples (gliding bird; diver vs. faller) |
| C3 | An open autonomous system that adaptively modulates its coupling generates all three conditions | moderate | §4 derivation from the definition |
| C4 | Minimal living organization is sufficient for agency; life is not necessary | moderate | §3 protocell and chemotaxis; §4 |
| C5 | Systems optimizing an externally fixed function are not models of agency | moderate as a consequence of C3 | §6, consequence (b) |
| C6 | Agency expands in complexity mainly through its spatial and temporal dimensions | weak | §5, declared less rigorous |
Concepts
- Individuality. The system’s own distinction of itself from its surroundings; used interchangeably with “identity”.
- Interactional asymmetry. The system’s capacity to modulate the constraints on its coupling at certain times.
- Normativity. Regulation against a condition that can be failed, with the condition set by the system’s organization.
- Autonomous organization. A precarious network of mutually dependent processes; captures “both the emergence of a self (autos) and that of norms (nomos)”.
- Sense-making. Interaction becoming significant for the agent because actions compensate deviations from a self-generated norm.
Connections
It builds on Varela’s autonomy, Jonas’s biological individuality, Ruiz-Mirazo & Moreno (2000) and Christensen & Hooker (2000), widening the latter two beyond metabolism. Sense-making is Di Paolo’s (2005) and Thompson’s, the notion De Jaegher & Di Paolo (LIT-571) extend to social interaction. Normativity is grounded with “Mossio et al. 2009” (LIT-570). Later, Bruineberg et al. (LIT-603) use the paper’s “interactional asymmetry” to name the arrow structure in a Friston blanket.
Bearing on the record
- Agent versus individual (ADR-024). This is the record’s cleanest source
for keeping
agencyandindividuationapart: individuality is necessary for agency and insufficient for it. - Markov-blanket individuation (LIT-526). The paper’s criteria imply that a partition read off statistical dependencies does not by itself make an individual, and C2 rejects correlational asymmetry. It does not discuss Friston, whose paper is four years later; the bearing is the record’s, not the authors’.
- THEORY filed as THEORY-071 (Proposed): minimal agency is an autonomous organization adaptively modulating its coupling for its own maintenance; individuality without asymmetry and self-generated norms is not agency.
- No instruction for machine-learning practice; consequence (b) is a remark about what counts as a model of agency, not a training method.
Limitations
- The authors call the modulation formalization incomplete: “what counts as a parametric modulation and what as a coupling between variables is often a matter of choice for the modeler”.
- How a system can be “sensitive” to norms that emerge from holistic dependencies is left open, as is the status of reflexes.
- No model is analysed; the definition is checked only against intuitive cases.
Open questions
- A modeller-independent criterion for interactional asymmetry, which the authors themselves ask for.
- Measures of individuality (closure, e.g. via category theory) and of normativity (homeodynamic stability) that would let the definition be applied “almost automatically” to a model, which is the test the authors set for their own definition.