NOTE-793: Abel et al. — Agency Is Frame-Dependent

Contribution

The paper states, as a thesis for reinforcement learning, that whether a system has agency depends on a reference frame: a set of commitments an observer makes before the question can be answered. It supports this by taking four conditions of agency in turn and citing, for each, a result or remark showing that the condition can come out either way under two plausible choices. Nothing is proved. The paper’s own words: “At present, our definitions, claims, and arguments are purely philosophical”.

Key insight

Each condition of agency hides a choice. Where is the boundary? Which causal variables are in the model? What makes a goal “meaningful” rather than an as-if description that fits every input-output system? Which changes in behaviour count as adaptation? Fix these and agency can be decided; leave them open and it cannot. The frame is not a defect to be removed but part of what a statement about agency says.

Assumptions

• A four-part account of agency, “due largely to Barandiaran et al. (2009)” (§1, Figure 1a): (1) individuality, a boundary separating the system from its surroundings; (2) source of action; (3) goals or norms that regulate interaction; (4) adaptivity, steering experience in light of the goals. The paper notes that Barandiaran et al. “combine normativity and adaptivity into one property”. It is a reconstruction, not Barandiaran’s three conditions (see Connections).
• Agency is binary on this account: a system has it if all four conditions hold. “Naturally there is room for developing a non-binary account.”
• Robustness to the definition is asserted, not shown: “regardless of how the semantics of ‘agent’ or ‘agency’ are worked out, the need for frame dependence will remain.”
• The setting is input-output systems as RL models them. The goal argument (Claim 3) is about inferring reward from behaviour.

Key results

There are no theorems. The four claims, as stated (§2):

• Claim 1 (individuality; adapted from Jiang 2019 and Harutyunyan 2020). “Nearly all agents admit many plausible boundaries one could draw that separates them from their environment. Moreover, key quantities of an agent can change depending on which boundary is chosen.” Support: Bohr’s stick, the extended mind, and Jiang’s Proposition 10, that the optimal policy and Bellman error of an RL agent depend on where the agent-environment boundary is drawn.
• Claim 2 (source of action; adapted from Kenton et al. 2023). “There exist cases with at least two plausible choices of causal variables where the former choice identifies an agent in the causal model, and the latter choice refutes the presence of an agent in the causal model.” Support: Kenton et al.’s own remark that discovering an agent “is relative to a frame – a choice of variables that appear in our causal model”.
• Claim 3 (normativity). “For nearly all cases of input-output systems, whether that system is meaningfully goal-directed depends on a reference point whose choice is arbitrary.” Support: every system can be described as if goal-directed (a broken thermostat that always sets 20° “has the goal” of 20°); in inverse RL the zero reward is consistent with every behaviour (Russell 1998; Ng and Russell 2000), and reward is underdetermined by behaviour without intervention. Occam priors and maximum entropy rule some goals out, but they are themselves extra principles.
• Claim 4 (adaptivity; adapted from Zadeh 1963 and Theorem 3.1 of Abel et al. 2023). “For many input-output systems, there will exist at least two reference frames (and in most cases, many more) such that according to the first reference frame the system is adaptive, while according to the second, the system is not.” Zadeh: “what matters is not whether [the system] is adaptive or not, but what … it is adaptive [to]”. A fixed policy is non-adaptive under one reference class of behaviour change and adaptive under another.
• The conclusion. “Since agency is simply the logical conjunction of these latter three properties conditioned on the choice of a boundary, then reaching a conclusion about agency itself must be made in reference to these commitments. In other words: agency is frame-dependent.”
• The reference frame is “a collection of these four commitments”: a boundary; “a reference object such as a set of causal variables”; “a principle for isolating whether the system is meaningfully pursuing a goal”; “a choice of what changes in behavior count as meaningful adaptation”. A boundary “could be a cut in a graph (Jiang, 2019) or a Markov blanket (Friston et al., 2009)”. A formal construction is left to future work.
• Choosing a frame (§3). “It is unclear which frame-selection principles are defensible”. Dennett’s intentional stance “could be formalised as a principle for selecting frames in terms of predictive or explanatory power”. Frames are likened to Marr’s levels and Dennett’s stances.

Claims

id	claim	strength	support
C1	Where the agent–environment boundary is drawn changes quantities such as the optimal policy and Bellman error	moderate	cited: Jiang (2019) Prop. 10, not reproduced
C2	Whether a causal model contains an agent depends on the chosen variables	moderate	cited: Kenton et al. (2023), LIT-909
C3	Goal-directedness inferred from behaviour needs an extra principle to rule out trivial goals	moderate	the inverse-RL underdetermination results, cited
C4	Whether a policy is adaptive depends on a chosen class of meaningful behaviour change	moderate	cited: Zadeh (1963); Abel et al. (2023) Thm 3.1
C5	Attributions of agency on the four-part account are relative to a frame	weak	the conjunction of C1–C4, informal
C6	The choices that make up a frame are arbitrary	weak	asserted; no argument that no choice is privileged
C7	Any basic science of agency requires frame-dependence	weak	asserted in the abstract and §3; not argued beyond C5

Concepts

• Agency. “A system’s capacity to steer outcomes toward a goal.”
• Reference frame (agent reference frame). The four commitments listed above, jointly sufficient to decide the four conditions.
• Frame-dependent. A property is frame-dependent when “reaching a conclusion about whether a given system possesses [it] requires an independent commitment whose choice is arbitrary”.
• Meaningful goal. Not defined. The paper’s point is that a definition requires a principle outside the system.

Connections

It is a short gathering of earlier results under one name. Claim 2 rests on Kenton et al. (LIT-909), whose frame-relativity the record already records in CLAIM-174. The principle it floats for choosing frames, predictive or explanatory power after Dennett (The Intentional Stance, LIT-430, held here unavailable; “True Believers”, LIT-398), is what Orseau et al. (LIT-911) formalise as an agent-versus-device comparison under a prior.

Its account of Barandiaran et al. is a reconstruction. Barandiaran, Di Paolo and Rohde (LIT-566, read in NOTE-458) give three conditions (individuality, interactional asymmetry, normativity), not four. Two of them are stated precisely so that no observer supplies them. The system must distinguish itself from its environment “without an observer”, or explanation regresses through observers. Its norms are set by its own organisation, and failure must be possible. Abel et al. argue that an observer’s boundary and an observer’s goal attribution are frame-relative. They do not engage the claim that an autonomous system draws its own boundary and sets its own norms. So against THEORY-071 the paper changes the subject rather than refuting it: it shows that behavioural or observer-side tests of these conditions are frame-relative, which Barandiaran et al. would grant.

Bearing on the record

• QUESTION-032. It supports the question’s framing: agency asked “relative to a level of description and a target”. It widens what the relativization must include. Beyond a level and a target, a frame fixes the causal variables, a principle for meaningful goals and a class of meaningful adaptations. It does not supply the test the question asks for, and gives no verdict on the thermostat (CASE-044): it uses the thermostat only to say that the verdict depends on the frame. It should carry no answers edge.
• CLAIM-174. Its “What it does not say” already concedes that test 2 inherits an observer-relative prior and test 3 Kenton et al.’s frame-dependence. This paper adds that test 3’s verdict also depends on which changes count as adaptation (Claim 4). Kenton et al.’s intervention on “how its actions affect the world” is one such choice, not the only one. It also predicts that test 1’s boundary, taken from coupling, is one more frame commitment, which is CLAIM-152’s caution in other words.
• CLAIM-194. It agrees that coupling criteria cannot by themselves supply individuality. It goes further, holding that no boundary is observer-free, which CLAIM-194 does not hold. CLAIM-194 instead locates the line at self-production, the move Abel et al. do not address.
• THEORY-232 (Watson, LIT-990). The two are close. Watson’s comparison of a whole’s agency with its parts’ is relative to a chosen decomposition and stipulated goals for the parts. Those are Abel et al.’s frame components (1) and (3). Watson concedes that relativity for one measure; Abel et al. argue that every account has it. Two differences:
  • Watson’s measure is comparative and stays meaningful within a frame. Abel et al. say nothing against comparison within a frame.
  • THEORY-232’s promote_when asks for an agency ranking stable across reasonable decompositions. Abel et al. predict no such stability in general, but they show no instance for Watson’s measure, so the paper cannot settle that theory either way.
• THEORY-071 and THEORY-208. Against Barandiaran’s absolute conditions, see Connections: the argument targets observer-side tests, not organisational closure. Against Floridi’s level-of-abstraction-free permissiveness, it supplies the relativity Floridi leaves unstated.
• LIT-969’s own filing text overstates it. It says the paper addresses “every essential property of agency in Barandiaran’s and Moreno’s accounts”. Moreno (2018) is cited as background and not analysed. And the paper takes no “direct position” on the thermostat, beyond saying that the verdict on it is frame-relative.
• No new THEORY in the record holds the thesis. It is filed from this reading as THEORY-249.
• No instruction for machine-learning practice; nothing for the anthology.

Limitations

• No definitions, no proof. The authors say so and list both as future work.
• Quantifiers drift. Claim 2 is existential (“there exist cases”), Claim 4 says “many”, Claim 3 “nearly all”. The conclusion is universal (“any measurement of a system’s agency must be made relative to a reference frame”). From existential premises it follows only that some attributions depend on the frame, not that all do. A system on which every plausible frame agrees is not ruled out.
• “Arbitrary” is asserted. That a choice is needed does not show that no choice is better. The paper itself floats predictive power as a principle of selection, which would make the choice principled.
• The goal argument is about inference from behaviour. It shows that behaviour underdetermines reward. It does not touch accounts on which goals are fixed by the system’s organisation (Barandiaran) or by intervention (Kenton et al. use interventions, and Amin et al. 2017, which the paper cites, recover reward with them).
• Each step is borrowed. The four supporting results are cited, not reproduced, and this reading did not check them.

Open questions

• A formal definition of a reference frame, and a proof that the four conditions are frame-dependent under it. The authors name this as the next step.
• Whether some systems are agents under every reasonable frame. A “frame-robust” core would make frame-dependence a matter of degree.
• Which frame-selection principle is defensible. Predictive power, as in Orseau et al., would turn the relativity into a model-comparison problem the record already has machinery for (CLAIM-174 test 2).