Agent-3 on the Horizon? Anticipatory Modelling of Aesthetic Processes in an Ecosystem of Human Actors and AI Agentic Systems

This study examines whether current developments in artificial intelligence are already establishing the functional, technological, and institutional conditions necessary for the anticipatory modelling of aesthetic processes distributed between human actors and AI agentic systems. It employs Agent-3—presented in the AI 2027 scenario as a hypothetical realisation of the superhuman coder (SC) milestone—as a scenario-based reference point. Agent-3 is treated neither as an existing system nor as an empirically substantiated prediction, but as a parametric model of the possible integration of capabilities that are already developing independently: multi-step reasoning, functional empathy, memory, tool use, recurrent agentic loops, implicit pattern-based inference, and limited self-improvement. The study compares these capabilities with the functional indicators of consciousness P1–P5 while maintaining a strict distinction between conscious access and phenomenal consciousness. Through the Model of Experimental Speculative Aesthetic Processes in AI Art (MEPAA), it proposes experimental and speculative protocols for investigating the causal participation of AI in aesthetic decision-making. It also considers the risk that the fundamental conditions of future aesthetic processes may be predetermined by a small number of foundation-model developers. The anticipated outcome is not evidence of AI consciousness or autonomous authorship, but the identification of conditions under which distributed aesthetic agency might emerge. The study argues that alternative aesthetic platforms should be developed before proto-generalist agentic systems become technologically and institutionally stabilised.

Agent-3 on the Horizon? Anticipatory Modelling of Aesthetic Processes in an Ecosystem of Human Actors and AI Agentic Systems

Proposal for a Preliminary Study

Agent-3 on the Horizon? Anticipatory Modelling of Aesthetic Processes in an Ecosystem of Human Actors and AI Agentic Systems

Abstract

Keywords: Agent-3, superhuman coder, pre-AGI, proto-generalist agentic regime, MEPAA, AAIA, AIARTWORLD, functional empathy, artificial intuition, aesthetic emergence, AI consciousness

Background and Definition of the Problem

Agent-3 does not currently exist as an empirically demonstrated system. Within the AI 2027 scenario, it denotes a hypothetical AI system corresponding to the superhuman coder (SC) milestone: a system capable of performing programming tasks relevant to AI research at the level of the most capable human experts, while operating considerably faster and through a large number of parallel instances (Kokotajlo et al., 2025). The AI 2027 document cannot therefore be employed as empirical evidence of the imminent arrival of artificial general intelligence (AGI), nor as a reliable chronological prediction. For the purposes of the present study, however, it may function as a scenario-based and parametric framework through which to model the implications of integrating capabilities that are already developing within separate AI systems.

Individual components anticipated by the Agent-3 scenario can no longer be regarded as entirely speculative. Reasoning models demonstrate multi-step problem-solving, the verification of intermediate results, adaptive shifts in strategy, and functionally defined forms of self-reflection. Research on DeepSeek-R1 has shown that some of these patterns can be developed through reinforcement learning without human-annotated reasoning trajectories (Guo et al., 2025). Nevertheless, verbalised reasoning chains cannot automatically be treated as transparent records of a model’s internal processes. Chain-of-thought outputs may produce persuasive retrospective rationalisations without identifying the actual causal factors that determined a response (Turpin et al., 2023).

Agentic architectures also employ feedback, episodic memory, and iterative cycles of reflection. The Reflexion system, for example, retains verbally formulated evaluations of previous failures and uses them to modify subsequent decision-making (Shinn et al., 2023). The Darwin Gödel Machine constitutes an experimental step towards self-improving agents, as it can iteratively modify its own code and empirically evaluate the success of those modifications. Its demonstrated capabilities, however, remain restricted to programming tasks, depend upon existing foundation models, and operate within a controlled environment subject to human oversight (Zhang et al., 2025). These results cannot therefore be interpreted as evidence of general or autonomous recursive self-improvement.

A further relevant domain is functional or computational empathy. Experimental studies indicate that AI can facilitate more empathic communication between humans and produce responses that evaluators perceive as empathic, understanding, or compassionate (Ovsyannikova et al., 2025; Sharma et al., 2023). Such findings provide evidence of a system’s ability to detect social and affective contexts and to generate pragmatically appropriate responses. They do not, however, demonstrate that AI experiences empathy as an affective or phenomenal state.

The concept of artificial intuition requires comparable caution. In this study, the term does not denote a subjective intuitive insight. Instead, it refers to a functional analogue of rapid, implicit, pattern-oriented inference through which a system selects a contextually relevant solution without explicitly deriving every intermediate step (Trovati et al., 2023). Artificial intuition defined in this way may be relevant to aesthetic decision-making because the selection of an aesthetic possibility does not depend solely upon the application of formal rules; it may also involve the recognition of relationships, similarities, deviations, and contextual relevance. Functional similarity, however, does not establish identity with human intuition or imply the existence of subjective experience.

The present state of research does not justify the claim that Agent-3 is immediately attainable. It does, however, support a more precise proposition: several functional components anticipated by the Agent-3 scenario already exist within distinct experimental systems and are progressively approaching the possibility of technological integration. This convergence provides a substantive reason to begin investigating future aesthetic processes before the emergence of a stable proto-generalist agentic system.

Central Claim and Research Urgency

The central claim of the study is that a distributed aesthetic process may emerge before the achievement of full AGI and without requiring the assumption of phenomenal AI consciousness. A sufficient condition may be a configuration in which an AI system acquires a stable and causally effective role in initiating, selecting, comparing, maintaining, and revising aesthetic possibilities. The determining question is therefore not whether AI ‘experiences beauty’, but whether it can co-determine the trajectory of an aesthetic process in a manner that cannot be reduced entirely to the execution of a single human instruction.

This question also possesses political and institutional urgency. If artistic research begins to examine future aesthetic processes only after advanced agentic systems have become stabilised, their fundamental conditions may already have been embedded within corporate architectures, training datasets, evaluation procedures, safety restrictions, and commercial models. It may therefore no longer be appropriate to wait for technology companies not only to construct future AI systems, but also to preconfigure the space of aesthetic possibility within which those systems will operate.

Experimental and speculative platforms for future aesthetic processes should consequently be developed before systems corresponding to the Agent-3 scenario become available. The objective is not to construct a ‘conscious AI artist’, but to establish alternative protocols and environments in which the distribution of aesthetic decision-making, memory, initiative, control, and responsibility among the human participant, the AI system, the technological protocol, and the environment can be investigated.

Pre-AGI as a Proto-Generalist Agentic Regime

The term pre-AGI is employed in this study as a provisional designation for a technological horizon situated between contemporary foundation models and hypothetical artificial general intelligence. Because AGI lacks a universally accepted scientific definition, the term proto-generalist agentic regime provides greater conceptual precision. It does not denote a weak form of already existing AGI. Rather, it describes a configuration combining multimodal processing, memory, longer-term planning, tool use, adaptive reasoning, the coordination of multiple processes, and partial functional autonomy.

Within this framework, Agent-3 serves as a scenario-based reference point rather than as a predicted technological event. The study does not test the accuracy of the AI 2027 scenario. Instead, it uses that scenario to formulate the question of what aesthetic consequences might follow from integrating presently separate capabilities into a single scalable agentic system. This use is consistent with the anticipatory and experimental-speculative orientation of MEPAA.

For the research project Aesthetics of Artificial Intelligence Art (AAIA; Marušiak, 2024), the proto-generalist horizon is relevant because an aesthetic process may become distributed before the emergence of artificial phenomenality. Within such a process, aesthetic decision-making is located neither exclusively within the human subject nor solely within the AI system. It develops through a temporally extended relational loop involving human intention, AI representations, memory, protocols, data, stochastic variation, environments, and institutional frameworks.

Functional Capabilities and Indicators of Consciousness

The methodological point of departure is the indicator-based approach developed by Butlin and colleagues (2023, 2026). Indicators P1–P5 are not treated as independent tests of consciousness, developmental stages, or variables that can be aggregated into a single score. Instead, they constitute heterogeneous sources of evidence derived from different scientific theories of consciousness. Their presence may increase the justification for a particular functional interpretation of a system, but it does not by itself demonstrate phenomenal consciousness.

Within the study, P1 denotes recurrent processing and temporal integration; P2, global availability and broadcast; P3, metarepresentation and self/other modelling; P4, attention and relevance selection; and P5, limited, auditable introspective reportability. The relationship between the capabilities under investigation and these indicators is summarised in Table 1.

Table 1
AI Functional Capabilities, Their Relationships to Indicators of Consciousness, and Their Interpretative Limitations

Functional capability	Significance for the aesthetic process	Possible relationship to the indicators	Unwarranted conclusion
Functional AI empathy	Recognition of social and affective contexts and generation of an appropriate response	Primarily P3: modelling the states of others	AI subjectively experiences empathy
Recurrent agentic loops	Iterative evaluation of outcomes, use of memory, and modification of procedures	P1: a partial functional analogue of recurrence and temporal integration	An external agentic loop is identical to biological or internal recurrence
Artificial intuition	Rapid pattern-based inference and selection of a contextually relevant possibility	Primarily P4: relevance selection	AI possesses subjective intuitive insight
Reasoning models	Multi-step problem-solving, verification, and adaptive modification of strategy	Partially related to P2 and P5	A verbalised chain of thought faithfully reproduces the internal process
Mechanistic global availability	Use of selected representations in subsequent reasoning and decision-making	P2: currently the strongest mechanistic evidence	A workspace-like organisation automatically demonstrates consciousness
Iterative self-improvement	Modification of procedures, tools, memory, or the system’s own code	Indirect relationship to P1, agency, and temporal continuity	The system develops autonomously and independently of humans, models, and infrastructure

Table 1 does not constitute a classification of conscious AI systems. It identifies possible functional analogies while simultaneously specifying the limits of their interpretation. The distinction between external recurrence in an agentic system and the internal recurrent processing postulated by certain theories of consciousness is particularly important. Repeated model calls, the use of external memory, and iterative revision do not establish that the architecture itself implements mechanisms comparable to biological recurrence.

Comparable caution is required when interpreting reasoning models. The ability to generate multi-step verbal explanations may improve performance and support functional monitoring of a solution. It does not, however, establish that the verbalised sequence constitutes an introspectively accurate record of internal computation. The assessment of P5 consequently requires behavioural evidence to be combined with causal interventions into internal representations.

Mechanistic Evidence and J-Space

Research in mechanistic interpretability is beginning to identify functional organisations relevant to conscious access. Gurnee et al. (2026) provide evidence of privileged J-space representations involved in verbal report, flexible reasoning, controlled modulation, and the causal redirection of outputs. This finding may be interpreted as the strongest current evidence for P2: the global availability of selected representations.

The result does not, however, demonstrate the existence of a complete global neuronal workspace or phenomenal consciousness. It remains unclear whether J-space produces a unified and continuous stream of representations, integrates sufficiently differentiated specialist processes, or maintains its functional organisation over time. The present study will therefore use mechanistic evidence not to confirm AI consciousness, but to determine whether a system’s internal representations can participate causally in aesthetic decision-making.

The causal criterion is central to MEPAA. An AI system cannot be regarded as a functionally significant component of distributed aesthetic agency merely because it generates aesthetically persuasive artefacts or employs the language of subjective evaluation. It must be demonstrated that particular internal or memory states influence the selection, maintenance, and revision of an aesthetic direction, and that interventions into those states produce predictable changes in the subsequent development of the process.

Research Aims and Questions

The study aims to define operationally a proto-generalist agentic regime relevant to AAIA and MEPAA, identify its partial realisations in contemporary AI systems, and develop experimental protocols for investigating distributed aesthetic processes. It will also examine how initiative, control, agency, interpretative power, and responsibility are distributed within these processes.

The principal research question is:

Under what functional, technological, aesthetic, and institutional conditions might a configuration approaching the scenario-based Agent-3 milestone support the emergence of aesthetic processes distributed between human actors and AI agentic systems, without that emergence being interpreted as evidence of phenomenal consciousness or autonomous aesthetic experience in AI?

The subsidiary questions concern which contemporary AI capabilities may be regarded as precursors of a proto-generalist agentic regime; how functional empathy, implicit pattern-based inference, multi-step reasoning, memory, tool use, and iterative improvement can be related to indicators P1–P5; when the generation of possibilities becomes causal participation in aesthetic decision-making; whether an aesthetic process can emerge without a stable self-model or phenomenal consciousness; how Human-Centred AI affects the distribution of control and responsibility; and which experimental findings would falsify the interpretation of distributed aesthetic agency.

MEPAA as a Research Protocol

MEPAA will be employed as an anticipatory research protocol rather than as a method for predicting the arrival of AGI. Its methodology will combine critical conceptual analysis, scenario analysis of AI 2027, a mechanistic case study of J-space, profiling of indicators P1–P5, and experimental-speculative modelling of aesthetic processes.

Each experimental scenario will specify which functions are performed by the human participant, the AI system, the protocol, and the environment; where the aesthetic objective originates; which component maintains and revises that objective; which decisions exert a demonstrable causal influence upon subsequent developments; and which observation would falsify the proposed interpretation. MEPAA will distinguish an empirically grounded vector derived from current research from a speculative what-if vector. The speculative component will not be presented as a prediction, but as a controlled artistic-research experiment involving alternative technological and institutional conditions.

The Political Economy of Aesthetic Emergence

Historical avant-gardes were formed through conflicts and interactions among artists, theorists, critics, institutions, and audiences (Bürger, 1984). Danto’s concept of the artworld explains how artistic status arises not solely from the properties of an artefact, but also from its theoretical, interpretative, and institutional environment (Danto, 1964). Danto’s work should not, however, be treated as a general historical theory of the development of the avant-garde.

Within AIARTWORLD, the conditions of aesthetic production may be determined to a considerably greater extent by the developers of foundation models. These developers make decisions concerning architecture, training datasets, post-training, safety policies, access conditions, memory, and the permitted extent of agentic autonomy. Although they do not create every individual artwork, they define the space of possibility within which aesthetic processes can occur. This produces a form of infrastructural or meta-authorial power that may intensify platform centralisation and the dependence of artistic practice upon technology providers (Burkhardt & Rieder, 2024; Zylinska, 2020).

This development is not entirely predetermined. Open-source models, local systems, artistic laboratories, and cultural institutions may redistribute some degree of technological power. Future aesthetic processes need therefore be neither wholly spontaneous nor entirely centrally controlled. More precisely, they may be emergent at the level of interaction while remaining infrastructurally conditioned and regulated at the level of models and platforms.

This argument does not entail the rejection of Human-Centred AI (Shneiderman, 2022). Rather, it rejects the uncritical assumption that the designation human-centred automatically guarantees meaningful human control. As memory, adaptability, planning, and tool-use capacities increase, control may shift from the direct management of individual operations towards the supervision of system-level objectives, constraints, and consequences.

Anticipated Conclusion

The study is unlikely to confirm the existence of Agent-3, artificial phenomenality, or an autonomous AI artist. It may, however, demonstrate that several functional components anticipated by the Agent-3 scenario are already emerging as distinct and experimentally investigable capabilities. Their future integration could create conditions in which an AI system ceases to function solely as a generative instrument and becomes a relatively stable and causally effective component of a distributed aesthetic process.

The anticipated outcome is therefore not the confirmation of a new artificial subjectivity, but the identification of an emerging functional space of aesthetic agency. Within this space, AI may participate in the initiation, selection, maintenance, and revision of aesthetic possibilities without being attributed subjective experience. Aesthetic agency would not appear as an intrinsic essence of the model, but as a relational effect produced through the coordination of the human participant, system representations, memory, protocols, and the environment.

The contribution of MEPAA will lie in its capacity to anticipate, model, and experimentally compare such configurations before proto-generalist agentic systems become technologically stabilised. Artistic research should not wait until the fundamental conditions of future aesthetics have been embedded within commercial platforms and subsequently accepted as immutable norms.

The time for artistic-research intervention does not begin after the arrival of Agent-3. It begins before it.

References

Bürger, P. (1984). Theory of the avant-garde (M. Shaw, Trans.). University of Minnesota Press. (Original work published 1974)

Burkhardt, S., & Rieder, B. (2024). Foundation models are platform models: Prompting and the political economy of AI. Big Data & Society, 11(2). https://doi.org/10.1177/20539517241247839

Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2023). Consciousness in artificial intelligence: Insights from the science of consciousness [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2308.08708

Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D. J., Constant, A., Deane, G., Elmoznino, E., Fleming, S. M., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., . . . VanRullen, R. (2026). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 30(6), 488–501. https://doi.org/10.1016/j.tics.2025.10.011

Danto, A. C. (1964). The artworld. The Journal of Philosophy, 61(19), 571–584. https://doi.org/10.2307/2022937

Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma, S., Bi, X., Zhang, X., Yu, X., Wu, Y., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., Liu, A., . . . Liang, W. (2025). DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature, 645, 633–638. https://doi.org/10.1038/s41586-025-09422-z

Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., Soligo, A., Bogdan, P., Ong, E., Wang, R., Thompson, B., Abrahams, D., Kantamneni, S., Ameisen, E., Batson, J., & Lindsey, J. (2026). Verbalizable representations form a global workspace in language models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2607.15495

Kokotajlo, D., Alexander, S., Larsen, T., Lifland, E., & Dean, R. (2025). AI 2027. AI Futures Project. https://ai-2027.com/

Marušiak, T. (2024, May 12). Research design AAIA 2024. Marussiac. https://www.marussiac.com/2024/05/12/c10/

Ovsyannikova, D., Oldemburgo de Mello, V., & Inzlicht, M. (2025). Third-party evaluators perceive AI as more compassionate than expert humans. Communications Psychology, 3, Article 4. https://doi.org/10.1038/s44271-024-00182-6

Sharma, A., Lin, I. W., Miner, A. S., Atkins, D. C., & Althoff, T. (2023). Human–AI collaboration enables more empathic conversations in text-based peer-to-peer mental health support. Nature Machine Intelligence, 5, 46–57. https://doi.org/10.1038/s42256-022-00593-2

Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., & Yao, S. (2023). Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36, 8634–8652.

Shneiderman, B. (2022). Human-centered AI. Oxford University Press.

Trovati, M., Teli, K., Polatidis, N., Cullen, U. A., & Bolton, S. (2023). Artificial intuition for automated decision-making. Applied Artificial Intelligence, 37(1), Article 2230749. https://doi.org/10.1080/08839514.2023.2230749

Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023). Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting. Advances in Neural Information Processing Systems, 36, 74952–74965.

Zhang, J., Hu, S., Lu, C., Lange, R., & Clune, J. (2025). Darwin Gödel machine: Open-ended evolution of self-improving agents [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2505.22954

Zylinska, J. (2020). AI art: Machine visions and warped dreams. Open Humanities Press. https://www.openhumanitiespress.org/books/titles/ai-art/