REVIEW 4 major objections 3 minor 26 references
Agency Is Frame-Dependent
T0 review · 4 major / 3 minor · reviewed 2026-08-09 · deepseek-v4-flash
Pith's one-line read Agency is not an intrinsic property of a system; it must be measured relative to a chosen reference frame.
desk verdict A useful synthesis of old ideas under a new label, but the central claim is under-supported because the paper never shows a single frame flips all four properties of agency at once. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is an 'agent reference frame,' defined as the tuple of commitments needed to measure agency: a boundary separating system from environment; a reference object, such as a choice of causal variables, for adjudicating the source of action; a principle for recognizing meaningful goal-pursuit; and a reference class of behavior changes that count as adaptation. The argument's load-bearing move is to show that for each of the four properties, at least two plausible frames yield opposite verdicts for the same system, so the conjunction—agency—is frame-dependent. The frame is an arbitrary upstream commitment that must be fixed before any agency measurement can be made.
What would settle it
A single system for which every admissible reference frame yields the same agency verdict would falsify the claim; for instance, if a formally defined thermostat could be shown to be non-agentic no matter which boundary, causal variables, goal principle, and adaptation class are chosen, the universal frame-dependence thesis would fail. A more direct test is to formalize reference frames and exhibit a system with provably frame-invariant agency status.
Extended reading notes
Core claim
In its own terms, the paper's central claim is that agency is frame-dependent: any measurement of a system's agency must be made relative to a reference frame. It takes the four-part account of agency—individuality, source of action, goal-directedness (normativity), and adaptivity—and shows that each part relies on an extraneous commitment that can be varied without changing the system. For individuality, multiple plausible boundaries exist; for source of action, the choice of causal variables can reveal or hide an agent; for normativity, behavior underdetermines goals; and for adaptivity, the choice of a reference class decides whether a policy is adaptive. Because agency is the conjunction of these properties, every agency verdict is relative to the frame. The paper is explicit that it offers a philosophical argument rather than a formal mathematical proof.
Load-bearing premise
The central argument assumes that the notion of a reference frame can be made precise enough to vary the four agency properties while keeping the system fixed, and that no frame is objectively privileged; if frames are unformalizable or constrained by objective standards, the claim becomes vacuous or false.
Editorial extensions
If this is right
- Empirical claims about agency must be reported together with the reference frame that produced them; a bare assertion that a system is or is not an agent is incomplete.
- Debates about whether particular systems—thermostats, robots, neural networks—are agents shift from a yes/no question to the question of which frame is being used and why.
- In reinforcement learning, choices about reward, goals, and policy adaptation carry implicit frame commitments that are currently left unspecified.
- A formal science of agency would require defining reference frames and then studying what remains invariant across them, rather than asking for a single intrinsic agency verdict.
- Plausible frame-selection principles, such as explanatory or predictive power, can make the intentional stance one method among many for choosing a frame.
Reading between the lines
- If the claim is right, then asking whether an AI system 'really' has agency is a category error: the real question is which reference frame is most useful or defensible for the purpose at hand.
- There may be an analogy with relativity: frame-dependent quantities coexist with objective invariants, and finding the invariants of agency could turn the philosophical claim into a productive research program.
- A concrete extension would be to formalize reference frames as tuples of mathematical objects and prove that simple systems (a thermostat, a neural network) admit frames assigning opposite agency verdicts; that would make frame-dependence a theorem rather than an argument.
- Because the paper identifies reward underdetermination as evidence for normativity frame-dependence, inverse reinforcement-learning methods that impose priors such as maximum entropy are implicitly choosing a frame—an insight that could make learned goal judgments more transparent.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper argues that agency is frame-dependent: any attribution of agency to a system is relative to a 'reference frame,' defined as a collection of commitments about boundaries, causal variables, goal-attribution principles, and reference classes for adaptation. It builds on a four-part account of agency from Barandiaran et al. (2009) and claims that each of the four properties is frame-dependent, citing prior results (Jiang 2019; Harutyunyan 2020; Kenton et al. 2023; Abel et al. 2023; Zadeh 1963). It concludes that agency itself is frame-dependent and discusses implications for RL. The paper explicitly stops short of formalizing reference frames or proving the main claim.
Significance. The paper is an original philosophical contribution that connects distinct strands of RL research to a long-standing question about agency. The individual claims are each grounded in published results, and the paper is honest about its limitations. If the central claim were established, it would have real consequences: agency would not be an intrinsic property of a system, and quantitative agency measurements would need to be indexed to frames. The novelty lies in the conjunction, but that conjunction is not currently proved; the paper's value at present is primarily as a research agenda rather than as an established result.
major comments (4)
- [Section 2 (summary paragraph)] The inference from per-property frame-dependence to frame-dependence of the conjunction is invalid as stated. For each property i the paper cites examples of frames f_i and g_i that flip property i in isolation, but it never shows that there exists a single pair of complete frames f and g—each containing a boundary, a causal-variable choice, a goal-attribution principle, and a reference class—such that under f the system satisfies all four properties and under g it fails to satisfy one or more. The examples use qualitatively different kinds of commitments for different properties: boundaries for individuality, causal variables for source of action, goal-attribution principles for normativity, and reference classes of behavior changes for adaptivity. The paper asserts that a frame 'must include' all four commitments, but this is the very statement that needs proof. Section 3 acknowledges that the formal proof is missing; this is not a peripheral gap but the load-bearing step of the argument.
- [Section 2, 'What is a Reference Frame?'] The definition of an 'agent reference frame' as 'a collection of these four commitments that allow us to determine whether a system has each of the four properties' makes the conclusion 'agency is frame-dependent' close to analytic: if any determination of agency requires such commitments, then agency is trivially relative to them. The nontrivial, falsifiable claim must be that there exist systems for which two different legitimate complete frames yield opposite agency verdicts, and that no principled criterion selects between the frames. The paper does not provide such an example, and it never defines what makes a frame 'valid' (it mentions 'many valid ways' but not the validity conditions). Without this, the reader cannot distinguish the intended substantive relativity from the trivial observation that all measurement requires a coordinate choice.
- [Section 2, Claims 1–4] The four claims are all 'adapted from' prior work, but the adaptations are not stated precisely. In particular, Claim 2 (adapted from Kenton et al., 2023) asserts that two choices of causal variables can identify or refute 'an agent,' yet Kenton et al.'s notion of agent-discovery may not coincide with the four-part Barandiaran et al. notion used here. Similarly, Claim 4 cites 'Theorem 3.1 of Abel et al., 2023' without stating the theorem or its conditions. Because the overall argument depends on these adaptations, the paper should state the original results and explicitly verify that they imply the frame-dependence claims in the sense required here.
- [Section 2, 'In summary' paragraph] The phrase 'the logical conjunction of these latter three properties conditioned on the choice of a boundary' is confusing. Earlier the paper lists four properties; if individuality is treated as a precondition that selects the boundary rather than as a property, this should be stated. More importantly, this sentence does not help with the composition problem: even if agency is a conjunction, the existence of separate flipping frames for each conjunct does not guarantee a single flipping frame for the conjunction. The paper also says at one point that the choice of reference frame is 'arbitrary' and at another that there are 'many valid ways' to formalize the components; these two statements are in tension and the paper never clarifies which one is intended.
minor comments (3)
- [Section 3] The word 'mostuseful' appears without a space in the paragraph on choosing a reference frame; it should read 'most useful.'
- [Reference list] The Ziebart et al. reference contains a typo: 'Artificiall Intelligence' should be 'Artificial Intelligence.' The Abel et al. reference also has an odd spacing in 'V . Roy.'
- [Throughout] The term 'frame-dependence' is used without a formal definition. A definition such as 'A property P is frame-dependent if there exist a system S and two frames f and g such that P(S) is true under f and false under g' would clarify what is being claimed and what would count as a counterexample.
Circularity Check
No significant circularity; the argument combines independent per-property results, though the informal reference-frame definition leaves an analyticity risk.
full rationale
The derivation chain is not circular: agency is treated as the conjunction of four properties, and each property's frame-dependence is supported by independent sources (e.g., Clark & Chalmers 1998 and Jiang 2019 for individuality; Zadeh 1963 and Abel et al. 2023 for adaptivity; Ng & Russell 2000 for normativity). The self-citations (Harutyunyan 2020; Kenton et al. 2023, which includes Richens; Abel et al. 2023, which includes several current authors) are real evidence: they are prior published or formal results rather than restatements of the present conclusion, and the affected claims also rest on non-overlapping citations. The main risk is not circularity but vacuousness: 'agent reference frame' is defined as the collection of commitments needed to determine the four agency properties, which makes 'agency determination depends on a reference frame' true by stipulation; the paper's substantive claim is that such commitments are arbitrary and can flip verdicts, which is argued by example rather than proven. The authors explicitly concede in Section 3 that they 'stop short of presenting a rigorous mathematical definition of reference frames, as well as a formal proof of the frame-dependence of agency,' and in Section 2 that their definitions and claims are 'purely philosophical.' That is an acknowledged rigor gap, not a hidden reduction of the conclusion to its inputs, so no circular step is identified.
Assumptions & free parameters
assumptions (4)
- domain assumption The four conditions of agency (individuality, source of action, normativity, adaptivity) are jointly sufficient for a system to be an agent.
- domain assumption Every input-output system can be interpreted as goal-directed (the 'as if' claim).
- standard math Frame-dependence of each of the four properties implies frame-dependence of their conjunction (agency).
- ad hoc to paper The choice of reference frame is arbitrary and no frame is privileged.
invented entities (1)
-
Agent reference frame
Cite this review
Pith. "Pith review of Agency Is Frame-Dependent." pith.science (2026). https://pith.science/paper/2TNVQ5CU
@misc{pith2026250204403,
author = {Pith},
title = {Pith review of: Agency Is Frame-Dependent},
year = {2026},
howpublished = {\url{https://pith.science/paper/2TNVQ5CU}},
note = {Machine review of arXiv:2502.04403}
}
read the original abstract
Agency is a system's capacity to steer outcomes toward a goal, and is a central topic of study across biology, philosophy, cognitive science, and artificial intelligence. Determining if a system exhibits agency is a notoriously difficult question: Dennett (1989), for instance, highlights the puzzle of determining which principles can decide whether a rock, a thermostat, or a robot each possess agency. We here address this puzzle from the viewpoint of reinforcement learning by arguing that agency is fundamentally frame-dependent: Any measurement of a system's agency must be made relative to a reference frame. We support this claim by presenting a philosophical argument that each of the essential properties of agency proposed by Barandiaran et al. (2009) and Moreno (2018) are themselves frame-dependent. We conclude that any basic science of agency requires frame-dependence, and discuss the implications of this claim for reinforcement learning.
Figures
Reference graph
Works this paper leans on
-
[1]
D. Abel, A. Barreto, B. V. Roy, D. Precup, H. van Hasselt, and S. Singh. A definition of continual reinforcement learning. In Advances in Neural Information Processing Systems, 2023
work page 2023
-
[2]
K. Amin, N. Jiang, and S. Singh. Repeated inverse reinforcement learning. In Advances in Neural Information Processing Systems, 2017
work page 2017
-
[3]
S. Armstrong and S. Mindermann. Occam's razor is insufficient to infer the preferences of irrational agents. In Advances in Neural Information Processing Systems, 2018
work page 2018
-
[4]
P. Ball. Organisms as agents of evolution. John Templeton Foundation: West Conshohocken, PA, USA, 2023
work page 2023
-
[5]
X. E. Barandiaran, E. Di Paolo, and M. Rohde. Defining agency: Individuality, normativity, asymmetry, and spatio-temporality in action. Adaptive Behavior, 17 0 (5): 0 367--386, 2009
work page 2009
-
[6]
H. Cao, S. Cohen, and L. Szpruch. Identifiability in inverse reinforcement learning. In Advances in Neural Information Processing Systems, 2021
work page 2021
-
[7]
A. Clark and D. Chalmers. The extended mind. Analysis, 58 0 (1): 0 7--19, 1998
work page 1998
-
[8]
D. C. Dennett. The intentional stance. MIT press, 1989
1989
Show all 26 references
-
[9]
F. I. Dretske. Machines, plants and animals: the origins of agency. Erkenntnis (1975-), 51 0 (1): 0 19--31, 1999
1975
-
[10]
Friedman
M. Friedman. Essays in positive economics. University of Chicago press, 1953
1953
-
[11]
K. J. Friston, J. Daunizeau, and S. J. Kiebel. Reinforcement learning or active inference? PloS One, 4 0 (7): 0 e6421, 2009
2009
-
[12]
Harutyunyan
A. Harutyunyan. What is an agent? http://anna.harutyunyan.net/wp-content/uploads/2020/09/What_is_an_agent.pdf, 2020
2020
-
[13]
N. Jiang. On value functions and the agent-environment boundary. arXiv preprint arXiv:1905.13341, 2019
1905 arXiv
-
[14]
Kenton, R
Z. Kenton, R. Kumar, S. Farquhar, J. Richens, M. MacDermott, and T. Everitt. Discovering agents. Artificial Intelligence, page 103963, 2023
2023
-
[15]
O. Klein. Glimpses of N iels B ohr as scientist and thinker. Niels Bohr. His life and work as seen by his friends and colleagues. London: Interscience Publishers, pages 74--93, 1967
1967
-
[16]
D. Marr. Vision: A computational investigation into the human representation and processing of visual information. MIT press, 2010
2010
-
[17]
A. Moreno. On minimal autonomous agency: natural and artificial. Complex Systems, 27 0 (3), 2018
2018
-
[18]
Moreno and A
A. Moreno and A. Etxeberria. Agency in natural and artificial systems. Artificial Life, 11 0 (1-2): 0 161--175, 2005
2005
-
[19]
A. Y. Ng and S. J. Russell. Algorithms for inverse reinforcement learning. In Proceedings of the International Conference on Machine Learning, 2000
2000
-
[20]
S. Russell. Learning agents for uncertain environments. In Proceedings of the Conference on Computational Learning Theory, 1998
1998
-
[21]
Tomasello
M. Tomasello. The evolution of agency: Behavioral organization from lizards to humans. MIT Press, 2022
2022
-
[22]
L. A. Zadeh. On the definition of adaptivity. Proceedings of the IEEE, 51 0 (3): 0 469--470, 1963
1963
-
[23]
B. D. Ziebart, A. L. Maas, J. A. Bagnell, A. K. Dey, et al. Maximum entropy inverse reinforcement learning. In Proceedings of the AAAI Conference on Artificiall Intelligence, 2008
2008
-
[24]
@esa (Ref
\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...
-
[25]
\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...
-
[26]
@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...
Reviewed August 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.