{"id":"48745c00-97d3-42de-aaca-dec7c20b2b60","arxiv_id":"2502.10420","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LMAs should be treated as systems whose agency is unstable and must be measured, not as normal agents with persistent identity.","lead":"This position paper argues that language model agents (LMAs) should not be treated as ordinary agents, because the language models beneath them are stateless, stochastic, sensitive to wording, and mediated by language. It proposes measuring 'agentic identity' before, during, and after deployment instead of assuming LMAs are stable, coherent actors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's categorical claim rests on the unproven Sec 5.3 conjecture that scaffolding cannot stabilize LMA identity; since an LMA is defined as LLM plus scaffolding, the paper must explain why external memory state cannot make the composite stateful.","rationale":"The reader's CONDITIONAL verdict is appropriate. The weakest assumption is indeed the claim that scaffolding can only mask, not cure, LLM pathologies. I would sharpen it: the paper never justifies identifying an LMA with its bare LLM rather than with the composite stateful system. Because Sec 1 and Sec 4 treat scaffolding as part of the LMA, the statelessness of the transformer does not by itself imply statelessness of the LMA. Memory is not 'just another form of context' from the system's perspective; it is mutable state that persists between turns. The paper's Sec 4.1 argument would also apply to any program whose behaviour is a pure function of its inputs plus a database, which we normally recognize as stateful. Similarly, stochasticity can be controlled by decoding parameters and verification wrappers; the paper notes temperature adjustments have 'mixed results' but does not quantify the residual instability. The Sec 5.3 trade-off conjecture is doing the categorical work and is undefended; the bias-variance analogy is suggestive but not a proof. This does not mean the paper's position is wrong: current LMAs deployed with naive scaffolding may indeed fail identity conditions, and measuring agentic identity is a reasonable proposal. But the strong, universal framing requires the conjecture, so a conditional verdict with the need for empirical or formal support is exactly right. No new objection changes the verdict.","tokens_in":22111,"tokens_out":5396,"duration_ms":53195,"concrete_test":"Take a standard scaffolded LMA (e.g., an LLM with append-only memory and a tool-use loop) and set decoding to temperature 0 with fixed seeds. Instrument the system to expose its full state S_t = (memory contents, conversation history, tool registry). Run the four identity conditions proposed in Sec 6.2 for N sessions with identical and reordered queries, measuring re-identification accuracy from S_t, contradiction rate, and goal persistence, and compare against a classical stateful agent (e.g., a database-backed rule-based agent). If the scaffolded LMA meets the identity conditions at the same level as the classical baseline, the Sec 5.3 conjecture is falsified for that configuration; if the paper responds that this configuration is too constrained to count as an LMA, it needs to say explicitly what minimal generative power defines the class.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest load-bearing step is the inference from LLM pathologies to LMA identity failure, made in Secs 3-4 and defended only by the conjecture in Sec 5.3. The paper defines an LMA as an 'agentic system based upon LLMs' and includes memory, tools, planning, and infrastructure in the architecture (Sec 1, Sec 4). The pathologies are properties of the bare LLM: statelessness (Sec 3.1), stochasticity (Sec 3.2), semantic sensitivity (Sec 3.3), and linguistic intermediation (Sec 3.4). When external memory is added, the paper says 'the LLM itself is never truly updated by these memory modules. Memory is just another form of context' (Sec 4.1). This is the pivotal move: it identifies the agent with the LLM rather than with the composite system. A classical agent also has a fixed transition function and mutable storage; its identity is anchored in the total state. If the memory store is part of the LMA and is updated between turns, the composite has a state that can be the basis for identifiability, continuity, persistence, and consistency. The paper asserts rather than proves that such state cannot satisfy these conditions. Its only support is the Sec 5.3 conjecture of a 'necessary trade-off between ontological stability ... and power', justified by analogy to bias-variance/no-free-lunch. The analogy is not an argument: deterministic systems can be universal, so determinism does not imply low power. Without a proof or evidence that stabilizing scaffolding must destroy the generative capabilities that define LLM-based agents, the categorical conclusion 'LMAs are not normal agents' rests on an unargued impossibility claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This Position paper argues that Language Model Agents (LMAs), defined as agentic systems built around LLMs and augmented with scaffolding such as memory, tools, planning modules, and infrastructure, should not be treated as normal agents. The authors identify four 'pathologies' intrinsic to LLMs: statelessness, stochasticity, semantic sensitivity, and linguistic intermediation. They argue that these pathologies destabilize the ontological identity conditions of agents—identifiability, continuity, persistence, and consistency—even when scaffolding is added. The paper concludes that scaffolding can at best mask these problems, conjecturing a necessary trade-off between ontological stability and generative power. The authors recommend that agentic identity be measured before, during, and after deployment rather than assumed.","tokens_in":22339,"tokens_out":5811,"duration_ms":49692,"significance":"If the central claim is accepted, the paper has significant implications for the design, governance, and trustworthiness of LMAs in high-stakes applications, potentially shifting the field away from treating LMAs as stable agents toward a more measurement-oriented approach. The paper's emphasis on operational evaluation of agentic identity is a constructive contribution, and its enumeration of LLM pathologies is a useful synthesis. However, the argument's categorical conclusion rests on an explicitly stated conjecture that is not proven, and the analysis does not fully address the composite nature of LMAs, which include external state. The paper is nonetheless a valuable position statement that identifies a genuine conceptual gap in current agentic AI discourse.","major_comments":[{"comment":"The paper's central conclusion that LMAs 'are not normal agents' (Sec 1.2) is load-bearing on the conjecture in Sec 5.3 that 'there is a necessary trade-off between the ontological stability of such systems on the one hand and their power on the other.' The only support offered is an analogy to no-free-lunch theorems and the bias-variance trade-off. This analogy does not establish the claim: deterministic systems can be universal (e.g., a deterministic Turing machine is Turing-complete), so determinism does not imply low expressive power. Moreover, the paper defines LMAs as including scaffolding (Sec 1, Sec 4), so the LMA is a composite system with external memory. The claim in Sec 4.1 that 'memory is just another form of context' identifies the agent with the LLM rather than the composite; a classical agent also has a fixed transition function and mutable storage, and its identity is anchored in the total state. The paper must either provide a proof or evidence for the trade-off, or restrict the claim to LLMs without scaffolding. As written, the categorical claim is not supported.","section":"5.3"},{"comment":"The four pathologies in Sec 3 (statelessness, stochasticity, semantic sensitivity, linguistic intermediation) are properties of the bare LLM, but the LMA is defined as an 'agentic system based upon LLMs' including memory, tools, planning, and infrastructure (Sec 1, Sec 4). The paper does not show that these pathologies persist in the composite in a way that defeats identity. For example, stochasticity can be bounded by constrained decoding or repetition penalties, semantic sensitivity can be reduced by input validation and structured tool schemas (Sec 4.2), and statelessness is directly addressed by external memory (Sec 4.1). The paper's response in Sec 5.3 that scaffolding only 'masks' problems is asserted, not demonstrated. Additionally, the identity conditions in Sec 2.2 are introduced as matters of degree (Sec 2.1 acknowledges agency 'is a matter of degree rather than a binary property'), yet the argument applies them as if any residual instability invalidates agency. The paper should state the threshold of stability required for 'normal' agency and explain why current scaffolding fails it.","section":"Secs 3-4"},{"comment":"The paper claims that 'regardless of which criteria of agency is chosen, LMAs struggle to satisfy it' (Sec 2.1), but it only analyzes its own four identity conditions (identifiability, continuity, persistence, consistency) and does not engage with alternative agency criteria such as behavioral or functionalist definitions, which might be satisfied by the composite system. The term 'normal agent' is never defined precisely; the argument oscillates between 'normal sense' (Sec 1.2), 'traditional agents' (Sec 3.1), and 'classical computational agents' (Sec 2.2.4). Without a precise target, the central claim is difficult to evaluate. The paper should either define 'normal agent' operationally or scope the claim to the specific identity conditions under discussion.","section":"Sec 2.1"}],"minor_comments":[{"comment":"The phrase 'persistence and and consistency' contains a duplicated 'and'.","section":"Abstract"},{"comment":"The phrase 'plethora if different queries' should read 'plethora of different queries'.","section":"Sec 3.1"},{"comment":"The phrase 'in analogous spirit with no free lunch theorems' should be 'in the spirit of no-free-lunch theorems'.","section":"Sec 5.3"},{"comment":"Several bibliographic URLs are truncated and will not resolve, including the Docker blog entry ending in 'llm-docker-for-local-' and the Fortune article ending in 'nvidia's-jensen-huang -'.","section":"References"},{"comment":"The sentence 'These characteristics of memory integration mean that proposed cognitive architectures such as COALA ... are problematised' is confusing because the preceding sentences discuss planning modules rather than memory integration; the referent of 'these characteristics' should be clarified.","section":"Sec 4.3"},{"comment":"In item (i) of Identifiability, 'criteria by which to identity an agent' should be 'criteria by which to identify an agent'.","section":"Sec 2.2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a useful position statement but the central claim is stronger than the evidence. The Sec 5.3 conjecture is explicitly labeled as a conjecture, and the argument would be more persuasive if scoped to current scaffolded LMAs or supported by a concrete analysis of why composite state cannot satisfy identity. The positive proposal for measurement is under-specified; even a preliminary operationalization would strengthen the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read this. First, the paper's useful core is the claim that LMA identity—identifiability, continuity, persistence, consistency—is a real problem and should be measured rather than assumed. That is a sensible, actionable message. Second, the paper's stronger claim that LMAs can never be normal agents is not actually argued. It rests on the Sec 5.3 conjecture that stabilizing scaffolding would destroy the generative power of LLMs, and that conjecture is asserted rather than defended.\n\nWhat is new here is the packaging: the four LLM pathologies (statelessness, stochasticity, semantic sensitivity, linguistic intermediation) are each well-known and well-cited, but organizing them as threats to ontological identity conditions for agents is a useful framing. The paper also does a fair job of engaging with counterarguments in Sec 5, and it explicitly acknowledges that scaffolding improves things. That honesty counts for something.\n\nNow the soft spots, in proportion. The gap you'll feel most is the move from 'the LLM is stateless' to 'the LMA is stateless.' The authors define an LMA as LLM plus scaffolding, and then in Sec 4.1 they say memory is 'just another form of context' and the LLM 'is never truly updated.' But a composite system with external memory does have state—the memory read and written between calls is part of the system. A classical agent also has a fixed transition function and mutable storage; nobody says it has no identity because its state is external to its transition function. The paper needs to argue why the composite state cannot anchor identity, and instead it offers an analogy to bias-variance that doesn't do the work. Deterministic systems can be universal, so determinism doesn't automatically impose a power cost. This is the load-bearing step, and it is unsupported.\n\nThe positive proposal—agent identity evaluations, measuring ontological robustness before/during/after deployment—is sensible but only sketched. No metrics, no methodology, no examples. For a position paper that might be okay, but it makes the paper more of a call to arms than a solution.\n\nOverall: if you want a clear statement of the conceptual worries about treating LLM agents as normal agents, this is a decent read. It deserves peer review because the question matters and the paper is serious, but a referee should push for either a weaker conclusion or a real argument against scaffolding. I'd probably cite it as a position statement, but I wouldn't rely on its impossibility claim.","headline":"Worth reading as a well-structured position paper, but the categorical claim overreaches: the key step that scaffolding cannot stabilize LMA identity is an unproven conjecture, so the paper is a call to arms, not a settled result.","tokens_in":22908,"tokens_out":2990,"would_cite":true,"duration_ms":27810,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"AI agents lack the stable identity real agency requires","keywords":["language model agents","agentic identity","LLM pathologies","ontological stability","scaffolding","agent evaluation","stochasticity","semantic sensitivity"],"falsifier":"A concrete falsifier: build an LMA with external memory, a stateful wrapper, and deterministic decoding, then run the same prompt sequence 10,000 times and adversarially perturb each prompt at the semantic level; if the system's outputs, goals, and self-description remain stable across trials and perturbations—showing measurable continuity, persistence, and consistency—while task performance matches or exceeds a baseline LLM, then the claim that scaffolding can only mask pathologies would be refuted.","tokens_in":21836,"feed_emoji":"🤖","tokens_out":5308,"duration_ms":44776,"temperature":0.7,"pith_summary":"This position paper argues that language-model agents (LMAs)—LLMs wrapped in memory, tools, and planning modules—should not be treated as ordinary agents with stable goals, boundaries, and identity. It claims that four intrinsic properties of the underlying LLM—statelessness, stochasticity, semantic sensitivity, and linguistic intermediation—destabilize the four ontological conditions that agency requires: identifiability, continuity, persistence, and consistency. Scaffolding can mask but not remove these pathologies, and the authors conjecture a necessary trade-off between a system's ontological stability and its generative power. The practical upshot is that LMAs' agentic properties must be measured before, during, and after deployment rather than assumed, because treating them as normal agents produces a false sense of utility and trustworthiness, especially in high-stakes settings.","feed_headline":"AI agents lack the stable identity real agency requires","feed_subtitle":"LLMs are stateless, stochastic, prompt-sensitive, and text-only; scaffolding masks, not cures, the instability.","key_machinery":"The analytic core is a pairing of four LLM pathologies with four ontological identity conditions. The pathologies are statelessness (no internal state transition between queries), stochasticity (sampling from token distributions), semantic sensitivity (small prompt changes producing large output changes), and linguistic intermediation (everything filtered through tokenized text). The identity conditions are identifiability (distinguishable from the environment and other agents), continuity (identity across time steps), persistence (identity across different instantiations and sessions), and consistency (non-contradictory description and action). The argument is that the pathologies violate or destabilize these conditions, and that scaffolding only relocates or masks the instability because it does not change the LLM's computation. A named conjecture completes the machinery: a necessary trade-off between ontological stability and power, analogous to bias-variance or no-free-lunch trade-offs.","core_discovery":"The paper's central claim is that LMAs are not agents in the normal sense, and that acting as if they are limits their usefulness. The mechanism is that four LLM pathologies—statelessness, stochasticity, semantic sensitivity, and linguistic intermediation—propagate upward and destabilize the ontological identity conditions (identifiability, continuity, persistence, consistency) that any agent must satisfy, regardless of which detailed theory of agency one adopts. Scaffolding such as external memory, tool use, planning modules, and infrastructure can improve surface coherence and provide audit trails, but because memory is only additional context ingested at query time and all action passes through text, the underlying pathologies remain. In response, the paper proposes agentic evaluations: mechanistic interpretability inside the model and agent-identity evaluations outside it, to quantify the degree to which a given LMA configuration preserves its ontology, so that residual instability is measured rather than assumed away.","pith_inferences":["If the trade-off conjecture is right, one can search for a Pareto front between identity-stability metrics and task-performance metrics across scaffolding configurations, predicting a negative correlation that is not an artifact of engineering quality.","The identity-crisis framing suggests a concrete evaluation target: an agent-identity score built from repeated-query consistency, cross-instantiation persistence, and adversarial semantic-perturbation sensitivity, which regulators could use to compare systems on a common scale.","The same logic extends to multi-agent systems: when LMAs share memory stores, tools, or the same underlying model, boundary blurring makes it ambiguous which agent caused which action, so accountability frameworks may need to track causal identity, not just output traces.","If linguistic intermediation is the irreducible bottleneck, grounding LMAs in formal, non-linguistic action schemas (typed tool contracts) could reduce semantic sensitivity without sacrificing generative breadth; this is a testable direction the paper does not explore."],"forward_implications":["Developers should stop designing LMAs on the assumption of stable identity and instead instrument systems to detect drift, contradiction, and boundary blurring during operation.","High-stakes deployments in finance, healthcare, and government should not rely on the apparent agentic coherence of LMAs without quantitative, risk-based evaluations of their identity robustness.","Scaffolding configurations should be evaluated not only on task performance but also on how well they preserve identifiability, continuity, persistence, and consistency; a configuration that genuinely stabilizes identity would be a real advance rather than a patch.","The proposed trade-off predicts that pushing LMAs toward full determinism and ontological rigidity will reduce their generative breadth, giving designers a measurable constraint for future agent architectures.","Measuring agentic properties before, during, and after deployment becomes a standard assurance practice, comparable to auditing, for any system claimed to be agentic."],"supporting_citations":[{"why":"defines what LMAs are and surveys their rise, setting up the object under critique.","marker":"(Xi et al., 2023)"},{"why":"supplies the standard textbook notion of an agent that the paper argues LMAs fail to satisfy.","marker":"(Russell & Norvig, 2021)"},{"why":"gives the classical agent properties of autonomy, reactivity, and proactivity that the paper claims are destabilized by LLM pathologies.","marker":"(Wooldridge & Jennings, 1995)"},{"why":"provides the transformer architecture and token-sampling mechanism underlying statelessness and stochasticity.","marker":"(Vaswani et al., 2017)"},{"why":"documents the illusion of state in sequence models, supporting the statelessness pathology.","marker":"(Merrill et al., 2024)"},{"why":"frames language models as stochastic parrots, grounding the stochasticity pathology and its risks.","marker":"(Bender et al., 2021)"},{"why":"introduces chain-of-thought prompting, an example of scaffolding that appears to add deliberation but remains exposed to the pathologies.","marker":"(Wei et al., 2023)"},{"why":"presents tool-learning with Toolformer, representing the tool-use scaffolding the paper argues cannot cure identity instability.","marker":"(Schick et al., 2023b)"},{"why":"proposes the COALA cognitive architecture, an example of planning and memory scaffolding that the paper says does not address the identity crisis.","marker":"(Sumers et al., 2023)"},{"why":"shows that optimality can depend on the choice of universal Turing machine, supporting the claim that linguistic intermediation is an additional abstraction layer that distorts agency.","marker":"(Leike & Hutter, 2015)"}],"fun_headline_variants":["LMAs aren't agents: stateless, stochastic, prompt-sensitive","Language model agents fail the identity test for true agency","Scaffolding can't fix that LLM agents lack stable identity","Stop treating LLM agents as if they were normal agents","Real agency needs continuity; LLMs are ontologically unstable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the conjecture that scaffolding—memory, tools, planning, infrastructure—can mitigate but never cure the LLM pathologies, because any cure would require giving up the generative flexibility that makes LLMs powerful; if some scaffolding configuration fully stabilized an LMA's identity while preserving power, the paper's central conclusion would collapse.","fun_headline_variants_meta":{"raw":{"variants":["LMAs aren't agents: stateless, stochastic, prompt-sensitive","Language model agents fail the identity test for true agency","Scaffolding can't fix that LLM agents lack stable identity","Stop treating LLM agents as if they were normal agents","Real agency needs continuity; LLMs are ontologically unstable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000553,"raw_usage":{"total_tokens":2628,"prompt_tokens":928,"completion_tokens":1700,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":1616}},"tokens_in":544,"tokens_out":1700,"duration_ms":11006,"temperature":1.0,"reasoning_tokens":1616,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:21:05.962262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete falsifier: build an LMA with external memory, a stateful wrapper, and deterministic decoding, then run the same prompt sequence 10,000 times and adversarially perturb each prompt at the semantic level; if the system's outputs, goals, and self-description remain stable across trials and perturbations—showing measurable continuity, persistence, and consistency—while task performance matches or exceeds a baseline LLM, then the claim that scaffolding can only mask pathologies would be refuted.","supporting_citations":[],"review_version":1}