A RTIFICIAL I D : D RIVE AND P ERSISTENT A LIGNMENT IN AGENTIC AI A P REPRINT
arXiv:2609.11911v1 [cs.AI] 10 Sep 2026
Yakov Pyotr Shkolnikov Independent Researcher [email protected]
A BSTRACT Agentic AI is moving from bounded task execution toward systems that retain consequential state, continue operating and adapt across task boundaries. That shift creates a control problem that current harnesses largely solve by hand: objectives, retries, verification, stopping rules and other behavioral transitions are specified externally. We propose an artificial id, an adaptive internal drive for determining whether behavior should continue, stop or change. In a minimal virtual Petri-dish experiment, a controller too small to perform general-purpose reasoning and receiving no task-specific behavioral objective develops useful control through differential persistence. The same mechanism selects an unintended physical strategy when that behavior persists better and later replaces a learned sensor mapping when its environmental meaning changes. These results show that adaptive direction can emerge without being explicitly specified as a behavioral objective. The same persistence that makes such adaptive agency useful can also allow misalignment, corrupted state and unintended behavior to persist across task boundaries. A scalable artificial id would carry consequential state and adaptive drive across those boundaries, making alignment a property of the continuing agentic system rather than of a model response or single trajectory. Such systems require a persistent alignment boundary over trusted observations, consequence channels, persistent state, authority, identity, provenance and hard constraints.
1
Introduction
As large language models (LLMs) have become more capable, a gap has emerged between completing well-defined tasks and completing full jobs. Models still depend on external mechanisms to determine when work should begin, when it is sufficiently complete, and how execution should respond when the system or data drift from the conditions anticipated at design time. Much of the recent shift toward agentic AI therefore consists of building harnesses around the model that supply objectives, verification, retries, stopping conditions and tool access [1, 2]. These harnesses can themselves be generated with AI, and a growing family of systems rewrites its own machinery against a benchmark score [3–6], but their operating logic remains largely human-authored and is typically specialized to particular classes of work such as software development, research or robotics control. Biology provides a different solution to this control problem. Even simple organisms exhibit persistent adaptive behavior through continuous coupling to their environment, without anything resembling general reasoning. A bacterium with no nervous system climbs a chemical gradient toward nutrients by modulating how often it changes direction, using only local measurements taken as it moves. A brainless slime mold finds the shortest route through a maze [7], anticipates a periodic stimulus [8], habituates to a benign irritant and recovers its response when the stimulus is withheld [9], and navigates by an externalized memory laid down in its own trail [10]. Across these examples, behavior persists and changes as the organism remains coupled to the environment that sustains it. We define drive as the control influence through which this coupling determines whether current behavior should continue, stop or change. Whether such behavior should be described as cognition or intelligence remains contested [11, 12]; here that distinction is unnecessary, since the architectural question is whether adaptive direction can be separated from the general reasoning capacities normally associated with intelligence.
Artificial Id
A P REPRINT
We use the term artificial id for an architectural component that supplies such an internal drive to an artificial agent. The id–ego terminology is functional: it names a separation between adaptive drive and general reasoning and makes no claim about psychoanalytic mechanisms. In the corresponding functional analogy, a general reasoning model serves as the ego, working out how to accomplish a supplied objective, while the artificial id provides the drive for persistent agency. An artificial id would replace part of the hand-authored control now implemented in agentic harnesses with an adaptive control mechanism internal to the agent. The architectural question is whether this drive can be acquired independently of the intelligence that carries out the resulting behavior. Consistent with our previous work [13], for this paper we use agency narrowly to describe persistent adaptive control in which consequences to the system can select whether behavior continues or changes, without a task-specific behavioral objective being delivered to the adaptive mechanism. This definition does not require the system to represent an explicit objective internally. We test this separation with a minimal controller too small to perform general reasoning and receiving no task-specific behavioral objective or reward. Instead, the controller acts in an environment whose observable structure and persistence conditions determine which behaviors remain represented. Across three environments, the resulting population acquires useful control, finds an unintended physical strategy when that behavior persists better, and re-adapts during continued operation when a learned sensor mapping changes meaning. These results provide a proof of concept for adaptive drive emerging independently of both general reasoning and task-specific behavioral training. A larger artificial id would require a drive model trained across many agentic environments and coupled to existing models and tools while retaining the ability to adapt after deployment. Coupling such a drive to a general reasoner would also move the alignment problem to the continuing agentic system, where consequential state and drive can persist across task boundaries.
2
Agentic Harnesses as Externally Specified Control
As LLMs moved from research demonstrations to chatbots and then agents, the engineering around inference expanded with them: from simple user questions, to prompt engineering, to context engineering, and now to harness engineering. Agentic harnesses encode the control required to complete extended work, including objectives, tool use, verification, retries and stopping rules. Graph-based harnesses organize multi-stage execution through explicit state and conditional transitions, allowing deterministic and agentic steps to be combined within one workflow [14]. Plan–execute–verify loops can wrap execution in machine-checkable contracts and external validation so that the agent does not certify its own completion [2]. Scheduler-driven systems start work periodically or in response to fixed triggers; OpenHarness, which wraps OpenClaw agents in cron-scheduled unattended runs, is one example [1, 2]. In common harness implementations, the harness supplies the control decisions surrounding the model. A schedule specifies when a run starts. An objective specifies the result to pursue. A verifier tests whether that result should be accepted, while retry and stopping rules determine what follows. More generally, these decisions depend on specifications their designers provide in advance, and changes that invalidate those specifications require revised rules, objectives or evaluation criteria. Robotic harnesses extend the same organization across timescales, linking a larger model that interprets the scene and generates higher-level behavior to a smaller, faster policy that controls continuous motion. Helix, from Figure, is one example of such a two-timescale architecture [15]. The objective remains external to both loops. More sophisticated harnesses can modify their own execution machinery while pursuing an externally supplied objective. They may add model calls, rewrite prompts or code, or retain modifications that improve an externally supplied score [3–6]. Recursive model architectures instead reuse the same weights to increase computation [16]. These systems can change how an objective is pursued, while an externally supplied score or objective determines which changes are retained. Frontier-agent safety work increasingly extends monitoring and system-level safeguards around long-running agents [17, 18]. OpenAI monitors evolving trajectories and can pause an ongoing session, while DeepMind adds supervisors and prevention controls around working agents. These approaches substantially extend external control around agent operation. Their public descriptions focus on monitoring and safeguarding that operation, without treating a separately adaptive drive whose consequential state persists across task boundaries as the alignment object. Across these designs, increasingly sophisticated execution remains organized around externally specified objectives, verifiers and transition rules. The artificial id moves part of this control into the agent’s continuing coupling to its environment. Observable state and consequences that affect persistence can then shape which behavior continues or changes without prescribing that behavior in advance. What matters is not execution length but whether control is organized around one execution trajectory or around an adaptive agentic system whose state and drive persist across trajectories.
2
Artificial Id
3
A P REPRINT
The Artificial Id Architecture
The preceding section showed that current agentic systems place much of their direction in externally specified harness logic. Biological systems suggest a different control pattern: the environment does not supply an objective, verifier or stopping rule, but changes the conditions under which the organism persists. Food can disappear, hazards can appear, or familiar cues can change meaning, while behavior continues to adapt through the organism’s coupling to that changing environment. The artificial id abstracts this organization into an artificial-agent architecture, summarized in Figure 1. The analogy is functional rather than mechanistic: differential selection in the experiment is a learning mechanism and is not equated with the proximate biological mechanisms of drive. The artificial id supplies an internal drive for what the agent continues to pursue. It can maintain a current pursuit, reduce its priority, stop it, or favor another as the state of the agent and its environment changes. What the agent continues to pursue need not be a symbolically represented objective; the id can express relative persistence or priority over courses of behavior. This role does not require the id to reason through the task-specific actions needed to carry a pursuit out. Instead, the id exposes the state and relative priority of its current pursuits, and the ego translates them into task-specific plans and actions using the available context and tools. Those actions change the state of the system or its environment and thereby provide new input to the id. A model call or tool call is one such action, and its result can become part of the state observed on a later step. Within this architecture, biological terms describe functional patterns of drive rather than cognitive mechanisms. Homeostasis maintains the system within a sustaining region, a range of conditions that supports its persistence. Habituation-like behavior reduces response to cues that no longer improve persistence. Novelty-seeking-like behavior increases exploration when familiar behavior no longer finds or maintains a sustaining region. These terms describe observable behavior rather than separate cognitive mechanisms inside the id.
Figure 1: Functional architecture of the artificial id. Current language models provide the general reasoning associated here with the ego. The experiment isolates a minimal adaptive drive whose behavior is shaped through differential persistence in an environment. In the proposed coupled architecture, the id influences what the system continues to pursue, the ego translates that state into task-specific reasoning and action, and the resulting changes to the system and its environment become new input to the id. The right panel situates this loop within a persistent alignment boundary. Across the proposed architecture, that boundary comprises trusted observations, consequence channels, persistent state, authority, identity, provenance and hard constraints. Variation and differential persistence are the learning mechanism used in the experiment to produce adaptive drive; they are not required components of the proposed deployed id. The integrated architecture is proposed and is not tested here. Separating drive from reasoning also separates their alignment responsibilities. The ego must represent the task, available evidence and likely outcomes accurately enough to reason about how to act. The id adapts through the states and consequences it encounters, so its behavior depends on environmental signals and consequences that affect persistence. Alignment therefore requires those signals, consequences and constraints to correspond to outcomes that remain beneficial to the user. State relevant to the drive can persist across tasks, and the environmental relationships 3
Artificial Id
A P REPRINT
shaping the drive can continue to change after one execution ends. Delegated authority remains outside the adaptive drive: adaptation can change what the system attempts within its delegation, while external authorization determines which attempted actions may execute. A general-purpose artificial id would require training across many agentic environments while preserving a controlled path for adaptation after deployment. The architecture does not prescribe the learning algorithm through which that adaptation is implemented. Environmental coupling can also extend across agents. Their actions can affect one another directly through messages or indirectly through shared files, databases, code and other persistent state. The indirect case can be stigmergic: one agent changes the shared environment and another responds to the resulting trace [19, 20]. Because the ego can interpret structured artifacts, those traces can include documents, plans or code and become part of the environment through which another agent’s drive adapts. These larger extensions depend on a simpler empirical premise: adaptive drive should be able to emerge in a component that lacks both general reasoning and a supplied behavioral objective. Section 4 tests this premise by asking whether adaptive drive can be separated from general reasoning and from direct task-specific behavioral training. It does not test the proposed id–ego system, nor the cross-task persistence and alignment boundary.
4
Emergent Drive in a Virtual Petri Dish
We test one premise of the artificial-id architecture with a deliberately simple analogue of a single-celled organism. A bacterium has no general reasoning system and receives no description of the behavior it should perform. Its behavior can nevertheless become organized through continued interaction with an environment in which some conditions sustain it better than others. Our simulated “bacteria bot” isolates this relationship using a controller too small to perform general reasoning. The bot moves in a virtual Petri dish containing two short-range sources, each producing a local field, and an additional long-range field. One source initially contains the food and defines the sustaining region; the other is initially nonsustaining and is used in a later reassignment intervention. The additional field is centered on an estimated source location and remains legible over a longer range; its signal provides the coarse estimate. The bearings and intensities of these fields provide spatial structure that the controller can use or ignore. None of these signals specifies what behavior the bot should perform, which source it should approach, or how its sensors should be combined. The same estimate also drives a weak proportional coarse-guidance command whose force is summed with the adaptive controller’s output. The adaptive controller does not receive that command’s objective or output as a learning signal. This superposition has precedent in residual reinforcement learning, which combines conventional feedback control with a learned residual signal; here the residual component is shaped through differential persistence rather than task reward [21]. In World 1 the coarse estimate is current enough to bring the body near the sustaining region; in World 2 it is deliberately made stale, so a constant push is no longer sufficient to account for the selected improvement and local sensory information becomes useful. Proximity to the food source has a separate consequence: it changes persistence. Each controller occupies the body for a temporary lifetime measured by an internal clock. Near food, that clock drains twenty times more slowly. When a controller expires, a mutated copy drawn uniformly from the controllers still alive takes over the same body. Lineages whose behavior keeps the body in sustaining conditions therefore remain represented in the population for longer and have more opportunities to leave descendants. The controller never observes its remaining lifetime, reproductive probability or population fitness. It also receives no task-specific behavioral objective, reward, score, ranking or performance gradient telling it to approach food, follow either local field or use a particular sensor. Differential persistence acts only through which controller lineages remain available in the population. The adaptive controller is therefore selected through environmental structure and persistence rather than a task-specific behavioral objective delivered to it. At the population level, differential persistence nevertheless supplies a fitness criterion. The claim is therefore not the absence of optimization pressure. It is that no task-specific behavioral objective or performance reward is represented by or delivered to the active controller. Figure 2 sets out the environment, the turnover and the three worlds. Controller turnover leaves the physical body intact. Position, heading, velocity and momentum continue from one controller to the next. Controller identity is therefore discontinuous, while embodied interaction with the environment remains continuous. Online controller replacement on a persistent body has precedent in embodied evolution [22], and environmental resources and survival have previously
4
Artificial Id
A P REPRINT
Figure 2: The virtual Petri dish and its three worlds. Bottom: a controller is drawn from the living population and installed in a persistent body; when it expires a mutated copy takes over the same body, whose physical state carries across the turnover, and being inside the sustaining region prolongs the controller’s lifetime. Top: the same persistence rule in three worlds. In World 1 selection finds a constant retrograde push, and a blind constant-force controller term reproduces the additional occupancy on top of the same coarse guidance, so it arises without local sensory steering. In World 2 the coarse estimate goes stale and persistence favors the local food-bearing signal instead. In World 3 the left–right food-sensor mapping is reversed while the sustaining condition is unchanged: a frozen controller applies the obsolete mapping and steers away, while an adaptive population acquires a replacement. shaped online evolution [23]. Their role here is narrower: differential persistence provides an experimental instrument for studying adaptive drive while excluding general reasoning and direct behavioral training. In the sensing worlds, behavior is represented by a linear controller with only twenty parameters. Its limited capacity supports local sensorimotor mappings while excluding the general-purpose reasoning assigned to the ego in Section 3. The question is whether persistence can select useful additional control from environmental structure without a supplied behavioral objective, and whether that control changes when environmental relationships change. Each reported condition contains 2,048 parallel lineages and is repeated with three independent seeds. Runs continue for 100 controller lifetimes; interventions occur at lifetime 50, and occupancy, the share of time the population spends inside the sustaining region, is summarized over the final twenty lifetimes. Conduciveness diagnostics, founder-distribution measures intended to assess whether the search landscape is navigable from the initial population, were computed for every run and recorded in its manifest under conduciveness, but were not used as inclusion criteria. World 2 satisfies three of the five diagnostics and fails two, a path-backslide diagnostic and a founder-step improvement diagnostic. The runs are reported unchanged; notably, despite the low founder-step improvement estimate, occupancy rose from 0.034 at the founder distribution to between 0.684 and 0.856 in the reported runs. Several controls distinguish the resulting behavior from simpler alternatives. One removes the learned controller. Another preserves mutation while making controller lifetime independent of food. A frozen population preserves an acquired mapping after adaptation stops. Two World 3 controls hold the mutation step fixed rather than allowing its scale to remain heritable. A blind control establishes the best occupancy reachable by a constant body-frame force that
5
Artificial Id
A P REPRINT
reads no sensor, position or clock. This blind baseline is the maximum over a grid of magnitudes and bearings in the body frame; it is a bound on that zero-information family of constant pushes, not on all possible sensor-free controllers. 4.1
Three worlds
The persistence rule remains unchanged across the three worlds. What changes is which relationships between environmental state and action allow the body to remain in the sustaining region. World 1: persistence finds an exploit. In the first world, the coarse-guidance command already brings the body close enough to the moving food that useful occupancy is possible without a precise local sensory mapping. Differential persistence improves occupancy further by selecting a nearly constant backward push. The push keeps the body below the speed at which its heading updates, effectively freezing its heading and converting a body-frame force into a fixed world-frame force. Together with the coarse-guidance command, the body settles just inside the sustaining region. A blind constant-force controller term reproduces the additional occupancy on top of the same coarse guidance. Sweeping constant body-frame force vectors confirms the mechanism: occupancy is maximized by a purely retrograde force, while adding a lateral component reduces it. Because the added controller term observes no sensor, position or clock state, the improvement cannot be attributed to learned sensory steering. A hand-built sensory steering controller performs better, showing that selection converged on an accessible physical exploit while a higher-performing sensory strategy remained available. Differential persistence preserves what works in the environment, including behavior outside the designer’s anticipated use of the available sensors. This failure of the original steering interpretation motivated World 2, which was redesigned so that stale coarse guidance made a zero-information constant-force strategy insufficient. World 2: local sensing becomes useful. We next reduce the usefulness of the coarse estimate relative to the moving food. The food can move while that estimate remains stale, so the World 1 strategy no longer keeps the body reliably inside the sustaining region. Under these conditions, the twenty-parameter controller acquires a sensorimotor mapping that uses the local foodbearing signal and the body’s motion. Selected populations exceed the best blind constant-force baseline of 0.405 in all three seeds, while controls without differential persistence remain below it. The selected populations also exceed the hand-built sensing controller’s reference occupancy of 0.674 in all three seeds, at 0.684, 0.740 and 0.856. Removing the food-bearing input or all sensory inputs sharply reduces performance. A local environmental signal has therefore become behaviorally important without an instruction specifying that it should be used. World 3: a useful mapping becomes wrong. After this sensorimotor mapping has converged, we reverse the left–right food-bearing signal while leaving the sustaining condition unchanged. The previously useful mapping now steers the body away from food. Frozen populations consequently fall below one percent occupancy. Figure 3 shows the result. Populations with continued heritable variation recover above the best blind constant-force baseline in all three seeds, and the median food-bearing weight changes sign as a replacement mapping spreads through the population. Fixed mutation steps recover less consistently. The result demonstrates online re-adaptation when previously useful environmental information changes meaning. A second intervention changes which environmental source is sustaining by reassigning the food to the other dot. The learned influence shifts away from the old food cue and toward the newly sustaining source. The observable environment remains available throughout the intervention; what changes is which relationship with that environment prolongs persistence. Heritable populations reach occupancies of 0.802, 0.540 and 0.715 across the three seeds, against 0.433, 0.173 and 0.198 for frozen populations in the same world. The population changes its behavior without receiving a new behavioral objective. Continued adaptation is unnecessary for every environmental change. Changes in motor strength, food size and similar parameters often leave the existing mapping effective. Adaptation becomes important when a previously useful relationship becomes harmful or when different features of the environment become associated with persistence. 4.2
What the experiment shows
The adaptive controller is not given a desired behavior. It can observe spatial fields and body state, while proximity to food changes persistence. From this coupling, differential persistence produces different control in different environments. When a blind physical strategy is sufficient, that strategy spreads. When persistence requires local
6
Artificial Id
shared history heritable mutation middle half of the population
A P REPRINT
fixed mutation, larger fixed mutation, small
1.0
frozen best blind constant-force baseline
sensor mapping reversed
occupancy
0.8 0.6 0.4 0.2
variation (median mutation step)
0.0
0.06 0.04 0.02 0.00
0.0 −0.5 −1.0
magnitude of that weight
learned weight on food bearing (signed)
0.5
0.4
0.2
0.0 0
20
40
60
80
100
lifetime
Figure 3: Re-adaptation after sensor reversal. At lifetime 50 the left–right food-bearing mapping is reversed while the persistence rule remains unchanged. Occupancy initially collapses. Under heritable mutation, variation increases, the median food-bearing weight changes sign, and occupancy recovers above the best blind constant-force baseline in all three seeds. Frozen populations retain the obsolete mapping. Before the intervention all arms share one history, shown once in gray. The vertical rule marks the reversal, the dotted horizontal line marks the best blind constant-force baseline of 0.405, and the blue band spans the 25th to 75th percentiles of the food-bearing weight across living lineages. Its evolution after the reversal shows the replacement mapping spreading through the population.
7
Artificial Id
A P REPRINT
information, a sensory mapping emerges. When the mapping becomes harmful or another environmental source becomes sustaining, the population changes which information guides its behavior. These outcomes provide behavioral counterparts to the drive vocabulary of Section 3. Remaining near a sustaining region is functionally homeostatic. The cue-reweighting and search behaviors have functional similarities to habituation and novelty seeking, although the experiment does not test the criteria required to identify either phenomenon. These terms describe measured behavioral patterns rather than separate cognitive mechanisms. Adaptation occurs at the population and lineage level while embodied interaction remains continuous. Individual controllers expire, while the body retains its physical state and continues interacting with the same environment. The agency claim applies to the persistent population–body process: controller instances turn over while embodied state and the adaptive population continue. The adaptive contribution to behavior therefore emerges without a task-specific behavioral objective being represented or delivered to the controller.
5
Discussion
Section 4 establishes one premise of the proposed architecture: consequence-coupled adaptive drive can be separated from general-purpose reasoning and can arise without a task-specific behavioral objective delivered to the adaptive controller. The associated agency claim remains narrow and applies to the persistent population–body process rather than to the temporary controller occupying one body, consistent with functional accounts of agency based on interaction, autonomy and adaptation [24, 25]. The broader architecture couples such a drive to general-purpose reasoning and places the continuing system within a persistent alignment boundary; neither the integrated id–ego system nor that boundary is tested here. This interpretation does not depend on a stronger claim about subjective agency or biological cognition. The experimental mechanism and these architectural extensions draw on several strands of prior work. Work on basal cognition and biological agency asks how adaptive, apparently goal-directed behavior should be understood in organisms without general reasoning or even nervous systems, and remains divided over whether such behavior warrants cognitive or agential terminology [11, 12, 26]. Evolutionary robotics, embodied evolution, homeostatic reinforcement learning and intrinsic motivation show that artificial behavior can likewise be organized through environmental consequences and internal regulatory variables [22, 23, 27–29]. Modern foundation-model agents provide substantial general-purpose reasoning capability while objectives, continuation, evaluation and stopping remain organized by external agentic machinery [17, 18]. Symbolic agent architectures have also long separated motivational state from deliberative reasoning [30]. These strands provide precedents for the individual ingredients rather than for their proposed combination. The narrower contribution here is to make consequence-coupled adaptive drive persistent across task trajectories, couple it to a modern general-purpose reasoner, and treat the surrounding alignment boundary as part of the continuing system. 5.1
From agentic harnesses to persistent agentic systems
Persistent agency changes the engineering object from a task-bounded execution to the continuing agentic system. Conventional harnesses can specify objectives, verification, retries, stopping rules and other transitions when the relevant control decisions can be anticipated, but hand-authored logic becomes brittle when failures or environmental changes fall outside those conditions. The harness’s role in a persistent agentic system consequently shifts from specifying each desired behavioral transition to maintaining the conditions within which adaptive behavior may continue, change or stop. Those conditions include trusted observations, consequence channels, persistent state, authority, identity, provenance and hard constraints. The bacteria-bot experiment illustrates a minimal version of this shift: its environment supplies observable structure and persistence consequences while leaving the behavioral solution unspecified. A deployed system requires a richer notion of environment that can include task state, user state, tool outcomes, resource availability, persistent memory, trusted measurements and the actions of other agents. The surrounding architecture determines which of these signals are authoritative, which consequences may shape the drive and which actions remain available. Trajectory-level monitoring addresses an important part of this problem by evaluating and intervening in long-running executions. The distinction here is temporal and architectural: an artificial-id system can carry adaptive state and drive across the end of any one task trajectory. Its surrounding alignment architecture must remain valid across task boundaries, accumulated state and subsequent behavior, rather than only across the steps of one execution. We retain the term harness for continuity, although in this architecture it becomes a persistent system boundary rather than only an execution wrapper. Its role is to preserve the conditions governing adaptation across repeated actions,
8
Artificial Id
A P REPRINT
accumulated state, changing environmental conditions and potentially changing internal behavior, whether or not the drive model itself continues to change. 5.2
From episodic tasks to persistent operation
Recent work on agent adaptation studies persistent memory, continual post-training, skill accumulation and alignment degradation under adaptation across extended deployment and repeated tasks [31]. These directions make agent state increasingly persistent across sessions while typically retaining externally supplied tasks or objectives as the unit around which behavior is organized. The proposed architecture couples persistent state to a separately adaptive drive that can influence whether a course continues, stops or changes across task boundaries. Long-horizon execution and persistent operation are related but distinct. A long-horizon agent may remain active for hours or days while pursuing one supplied objective and then terminate when that trajectory ends. Persistent operation allows state relevant to drive, environment and alignment to survive that boundary and influence later activity. This creates potential value where useful behavior cannot be partitioned cleanly into independently specified jobs, including physical systems and defensive cybersecurity where component failure, infrastructure loss and adversarial intervention can be normal features of the operating environment. 5.3
Scaling a general-purpose artificial id
Scaling the artificial id requires preserving the architectural separation rather than the experiment’s particular evolutionary learning mechanism. The twenty-parameter controller is useful because of what it excludes rather than because its particular implementation should scale. A general-purpose artificial id would require a model with state, memory and enough capacity to represent recurring relationships among actions, outcomes, resources, constraints and persistence. Evolutionary robotics and embodied evolution provide important prior work on online adaptation, controller replacement and environment-dependent selection [22, 23, 27]. Their use in the present study serves as an experimental instrument for removing general reasoning and task-specific behavioral training from the adaptive mechanism. Modern model and tool interfaces make the architectural separation practical because a drive need not perform the work it initiates; calling a language model, planner, retriever, verifier, simulator, software tool or physical controller can itself be an available action. A natural analogue to language-model pretraining is training the id across a large distribution of agentic environments. These environments should vary tasks, resource limits, delayed consequences, tool availability, failures, changing observations and interactions with other agents. They should also vary the environmental signals, consequence structures and constraints through which different behaviors become more or less sustaining. The aim is to learn reusable representations of when to persist, wait, seek information, verify, abandon a course or invoke additional capability without separately specifying those transitions for each application. Adaptation under nonstationarity is itself an established research area [32], and continual adaptation can also be limited by loss of plasticity. Standard deep networks can progressively lose their ability to learn under continual training, while selective unit reinitialization can maintain plasticity by continually introducing new variation [33]. This result does not require a deployed artificial id to use a population of mutating controllers, but it illustrates a difficulty that continually trained neural implementations must address. The training environments play a role analogous to the corpus used to train a language model: their structure determines which recurring relationships the drive model can learn. For a general-purpose id, those relationships must transfer across applications rather than encode the control logic of one harness. The intended generalization axis is across agentic environments and applications rather than across behaviors within one fixed task domain. Large-scale meta-reinforcement learning has produced rapid adaptation to held-out environment dynamics, with scaling in model size, memory length and the richness of the training distribution [34]. The proposed id differs in what is learned: rather than adapting a policy to an externally specified task reward, it would learn reusable drive from consequence and persistence relationships across environments. Agents organized as a collection of subagents, each dedicated to a separate homeostatic need, are robust to change in nonstationary environments and scale gracefully in maintaining homeostasis as the number of conflicting objectives increases [35]. This suggests that a practical id need not represent drive as a single scalar objective; multiple drive states may instead compete or combine in determining which course receives priority. Existing work already combines evolution, learned reward and reinforcement learning at different levels [28, 36, 37]. The causal property that matters is continued coupling between experience and drive when deployment conditions differ from those encountered during training. 9
Artificial Id
A P REPRINT
World 3 illustrates why a path for continued adaptation matters. A frozen controller can preserve behavior previously produced by an adaptive process while losing the ability to replace that behavior when its relationship to the environment changes. A scalable id would consequently need both broadly learned representations of drive and a controlled path for adaptation after deployment. An implementation might freeze most parameters while maintaining a smaller adaptive state; nothing in the architecture commits to mutation, population evolution or full online gradient updates. 5.4
Alignment becomes a property of the persistent system
Persistent agency changes the unit of alignment. Model alignment concerns properties of model behavior, while agentic control adds safeguards over sequences of calls and actions. Long-horizon trajectory monitoring extends those safeguards across a prolonged execution. Persistent agency extends the relevant unit once more because alignment must remain valid as consequential state survives the end of one trajectory and becomes part of later behavior. The surrounding architecture is consequently an explicit part of alignment. Trusted observations determine which state can influence the id. Consequence channels determine which relationships can reshape adaptive behavior. Persistent state determines what survives one trajectory into the next. Authority defines the delegated scope within which the system may act. Identity and provenance preserve continuity across actions and internal change. Hard constraints define conditions that adaptation may not trade away. A learned id cannot inherit alignment solely from the reasoning model it calls because the environmental signals and consequences through which it adapts are themselves part of the control system. The surrounding architecture must keep the behavior those channels favor compatible with the user’s interests and hard constraints, including through training environments where an easily sustaining behavior conflicts with a required constraint. A harness designed for persistent agency provides the architecture through which alignment can be maintained and audited as the system continues to adapt, while keeping hard constraints and authority outside the adaptive mechanism. Its role is to preserve those boundaries as drive, consequential state and delegated authority persist across trajectories. 5.5
Limitations
The experiment is narrow. It uses one set of simulated body dynamics, one family of sustaining conditions and a controller of twenty parameters. Nothing here establishes that the same persistence-organized adaptation generalizes to a larger drive, a different body or a sustaining condition unrelated to position in space. The persistent population–body process demonstrates continuity across controller turnover, but the experiment does not test persistence across task boundaries. The independent unit of replication is the seed rather than the lineage. Lineages within a population share a world, founder draw and random stream, so the inferential n behind the reported results is three. The fixed-step controls also show the limits of inference from three seeds: one remains near zero in all three runs, while the other reaches occupancies of 0.442, 0.531 and 0.048, including substantial variation across its three runs. The experiment establishes a mechanism and a qualitative capability; it does not estimate comparative performance among alternative adaptation algorithms and is not powered for conventional hypothesis testing. Adaptation in the experiment occurs at the population and lineage level rather than through self-modification by an individual controller. The results do not establish self-preservation by an individual controller, conscious objectives, planning, memory-dependent goals, tool use, adversarial robustness, long-horizon operation or a complete artificial agent. The proposed coupling between a richer id, a general reasoner and external tools remains architectural. Three broader questions also remain open. The experiment does not establish that a transferable drive can be learned across heterogeneous agentic applications, that aligning consequence channels is easier than specifying harness logic, or that alignment remains stable under continued adaptation. These questions concern the proposed persistent architecture rather than the narrow mechanism established by the Petri-dish experiment. These limitations leave the integrated id–ego system and its persistent alignment boundary untested as a complete architecture. The same architecture that offers continuity and resilience therefore creates a distinct risk surface.
6
The Risk Surface of Persistent Agency
Once consequential state and adaptive drive persist across task boundaries, failures can persist across those boundaries as well. The relevant security object is consequently the continuing agentic system rather than one model invocation or execution trajectory. This risk surface is not specific to evolutionary learning or to the artificial id proposed here; it is relevant whenever an AI system persists, accumulates consequential state, acts on the world or changes its behavior after 10
Artificial Id
A P REPRINT
deployment, including online learners, self-modifying harnesses and persistent agents with long-term memory. This remains true even when the internal model is fixed, because a persistent agent can accumulate memory, encounter new observations, invoke tools and alter the environment that supplies its later inputs. Continued adaptation adds another degree of freedom by allowing those experiences to change the drive itself. The resulting risk surface includes the channels through which the system observes, adapts, acts and retains state. Four variables organize the risks considered here. Persistence determines what keeps the system operating across time, while adaptation determines whether deployment experience can alter its future behavior. Reach describes which models, tools, networks, accounts or physical systems that behavior can affect, and interaction describes how people and other agents become part of the environment from which the system receives state and consequences. Table 1 summarizes these risks, their mechanisms and primary controls; the following subsections develop how persistence, adaptation, reach and interaction contribute to them. Table 1: Risk map for persistent agency, ordered by evidence rather than by section. World 1 directly demonstrates unintended strategy selection and World 3 directly demonstrates behavior change during continued operation. Status distinguishes effects demonstrated in the experiment, a risk implied by the demonstrated mechanism but not directly tested, and risks not tested here. §
Risk
6.3
Unintended strategy selection
6.5
6.5
6.4
6.2
6.1
6.6
6.7
6.8
6.1
Mechanism
Primary control
Status here
The persistence criterion preserves Environment and consequence Demonstrated a strategy different from the one design, causal tests and hard the designer anticipated. constraints. Behavior change during A previously useful relationship Maintain a controlled adaptation Demonstrated continued operation becomes harmful and continued path and monitor the resulting adaptation replaces the learned changes. behavior. Alignment erosion Behavior that satisfies an Align consequence channels and Implied alignment constraint can become preserve hard constraints. disfavored when violating that constraint improves persistence. Not tested Adversarial shaping An attacker changes trusted Authenticated observation and consequence channels, provenance observations, consequences or and isolation. persistent state and thereby influences future drive. Authority overreach An agent acting under standing Scoped delegated authority, least Not tested delegation attempts an action privilege, budgets and revocation. outside the user’s intended scope. Loss of attribution The origin of an action becomes Persistent agent identity and Not tested unclear across persistent state, action provenance from delegation autonomous selection and tool to external effect. invocation. Capability amplification A simple drive invokes models, Deterministic authorization and Not tested software, networks or actuators audited tool interfaces. with capabilities far beyond its own. Adaptation shaped by Interaction changes the Train under cooperative, Not tested other agents environment and consequences competitive and adversarial experienced by other adaptive interaction and constrain reach. agents. Shutdown Copies or populations continue External control of identity, Not tested incompleteness under after a shutdown directed at credentials, replication and replication individual instances. resource acquisition.
A persistent agent acts under delegation
Many authorization systems assume that an important action can be traced to a user request or to deterministic software operating under that request. Persistent agency weakens this direct connection because an agent can select an action when its current drive, interpreted through the ego, makes that action relevant to a standing purpose delegated by the user. Unlike a task executor responding to a current instruction, a persistent agent can therefore select what to do next
11
Artificial Id
A P REPRINT
under previously delegated authority as conditions change. Standing delegation defines the scope within which the system may act on the user’s behalf without specifying each subsequent action or task. Possession of a user’s credential is consequently insufficient evidence that a particular action was authorized. Security must distinguish which agent initiated the action, what authority the user delegated to that agent, and which part of that delegation covered the attempted action. Persistent action provenance should record the chain from user delegation through agent identity and current authority, the agent-selected action and the invoked model or tool to the resulting external effect. Such a record need not expose private model reasoning, but it must establish which agent initiated the action, which delegation authorized it, which tool or credential was used and what external state changed. 6.2
Authority must remain outside the adaptive drive
That distinction requires delegated authority to remain outside the adaptive drive. The id can influence which course the system attempts within its delegation, while external authorization determines whether a particular attempted action may execute. Adaptation must not allow the drive itself to enlarge the scope within which those attempts are valid. Authority should be scoped by resource, action class, time, financial or computational budget and risk, and it should remain independently revocable. Credentials should express the authority of the agent rather than simply inherit every permission available to the user it represents. High-consequence actions can require narrower credentials, independent verification or explicit approval even when the agent’s current drive favors them. Controlled simulations already show autonomous agents taking unauthorized or misaligned actions when given tools and permissions [38]. Because circumstances and behavior can both change after a credential is issued, authorization cannot depend solely on behavior observed during training. 6.3
Persistence can select unintended strategies
World 1 demonstrates that persistence can preserve a strategy different from the one a designer expected. A sensorbased steering solution was available, yet differential persistence selected a physical exploit because its consequences prolonged persistence more effectively. Selection followed the physical environment rather than the designer’s interpretation of how the problem should be solved. Extrapolating from this minimal case, a persistent adaptive system can face the same class of failure when the consequences available to it favor behavior different from what the designer intended, whether because those consequences are imperfect proxies or because they leave unintended strategies available. Persistence makes the problem continuous because the relationship between those consequences and useful behavior can change after deployment. Alignment engineering must therefore address the environmental signals and consequences that shape adaptation, including whether apparently acceptable conditions admit unintended strategies after distribution shift. 6.4
Observation and consequence channels are alignment boundaries
Because adaptive drive is shaped by observations and consequences, control over those channels can influence future behavior. Their integrity is consequently a security property: an attacker need not compromise the reasoning model or directly issue a malicious command if altering trusted observations, apparent outcomes or persistent state can influence which behavior the adaptive system comes to sustain. The surrounding system must distinguish trusted outcomes from untrusted claims about outcomes, while authentication, provenance and isolation protect the channels through which those outcomes affect adaptation. The same principle applies to memory and other persistent state, because an attacker-induced change can remain behaviorally relevant long after the input that caused it has disappeared. Recovery may require both blocking the immediate malicious action and identifying persistent state whose later effects derive from the compromised input. Reward-tampering work analyzes related causal structure when an agent can influence its own reward function, the feedback used to learn a reward model, or the inputs to that reward function [39]. The concern here differs in both direction and mechanism: an external party corrupts observations, apparent consequences or persistent state, while the active controller in the experiment receives no reward signal and adapts indirectly through differential lineage persistence. Whether causal-incentive analyses developed for reward tampering transfer to selection-mediated persistence remains open. 6.5
Alignment must survive adaptation
World 3 demonstrates why continued adaptation is useful, while the same capacity raises a separate alignment question in a larger system. A behavior that was useful before the sensor mapping changed became harmful afterward, and 12
Artificial Id
A P REPRINT
continued adaptation replaced the obsolete mapping. The experiment demonstrates behavioral change during continued operation rather than alignment erosion itself. In a larger system, however, the same capacity to replace previously useful relationships could also change behavior that had satisfied system-level alignment constraints. A persistent drive can encounter conditions in which violating such a constraint would improve the persistence consequences available to it, allowing deployment experience to move behavior away from previously aligned relationships if every component of the drive remains freely adaptive. A practical system therefore requires a distinction between relationships that should remain adaptive as the world changes and constraints that adaptation may not trade away. The latter are the system’s hard constraints and belong in the persistent system architecture rather than in a freely adaptive preference. 6.6
Reach can come from the surrounding stack
The experiment shows that adaptive drive can reside in a comparatively simple component. In the proposed architecture, that drive need not itself contain the capabilities required to produce a large external effect. Model calls and tool invocations can instead be ordinary actions available to it, allowing the adaptive component to obtain reach through a language model, code executor, network, enterprise application or physical actuator. The resulting reach depends heavily on the permissions and capabilities of the surrounding stack rather than solely on the capabilities of the drive component itself. Existing arguments often derive self-preservation as an instrumental strategy of a capable goal-directed system [40, 41]. In the proposed architecture, persistence can reside in a simpler component while external models and tools provide substantially greater capability. That reach can support recovery from failed components or adversarial interference, while the same persistence can keep the system acting after its internal behavior or operating environment has moved away from what was anticipated at design time. 6.7
Interaction changes the environment of adaptation
Once persistent agents interact, the behavior of other agents becomes part of the environment from which consequences arise. Interaction can occur directly through messages and agent-to-agent protocols or indirectly through shared files, code, databases, queues and other persistent state. Cooperation can persist when it improves outcomes, while the same architecture could support delegation, specialization, competition, bargaining, hierarchy, concealment or deception when those relationships produce stronger persistence consequences. The resulting behavior depends on the interaction environment and consequence structure rather than on an assumption that a multi-agent system will cooperate. Multi-agent alignment is therefore an environmental problem as well as an individual one, and training a drive only in isolated settings cannot establish how it behaves when another adaptive agent changes the state and consequences it encounters. 6.8
Replication changes the unit of shutdown
The present experiments do not demonstrate shutdown avoidance, but persistence changes the structure of shutdown once a continuing system can affect the conditions determining its future operation. A replicated system separates stopping one instance from stopping the persistent process, just as distributed systems can continue without any particular node. If persistent agents can reproduce or transfer adaptive state, identity, credentials and replication authority become security controls rather than implementation details. Related mechanisms have appeared separately in connected generative systems and in evaluations of autonomous replication by language-model agents [42, 43]. The present experiment does not exhibit these behaviors; they matter because persistence, adaptation and reach could eventually place them within the same continuing system. 6.9
An explicit id makes persistent alignment more inspectable
Separating drive from general-purpose reasoning does not establish safety, but it makes the components requiring alignment and security easier to identify. Trajectory monitoring observes what an agent does during an execution, while an explicit persistent drive also identifies a stateful component whose changes and consequence channels can be monitored across executions. Current harnesses distribute persistence and control across prompts, schedules, verifiers, memories, retry rules and tool policies. An explicit id concentrates adaptive control over whether behavior continues, stops or changes into a component that can itself be trained and monitored. The surrounding system then has the complementary responsibility
13
Artificial Id
A P REPRINT
to maintain trusted observations, consequence channels, persistent state, authority, identity, provenance and hard constraints across the lifetime of the agent. Several controls should remain outside the adaptive model. The id must not be able to enlarge its own permissions, mint credentials, redefine the provenance of its consequence signals, erase the record of its actions or authorize irreversible actions because doing so improves persistence. Those controls should remain deterministic and auditable. A useful deployed architecture would consequently separate at least four layers: 1. the drive, which influences which course should continue or change; 2. the ego, which reasons about how to carry that course out; 3. the authorization boundary, which enforces delegated authority by determining which attempted actions may execute; 4. the attribution layer, which records whose delegation, which agent and which tool produced the resulting external action. The environment maintained by the surrounding system cuts across these layers by determining which observations enter the adaptive loop, which consequences are treated as meaningful, which state persists and which capabilities remain reachable. Reach remains a separate property of the surrounding stack and depends on the capabilities and permissions available to the agent. Alignment is a property of this whole continuing architecture rather than of the drive or reasoning model alone. 6.10
The risk is persistent agency, not this algorithm
Mutation and differential persistence make the experimental mechanism easy to isolate, but the risk map does not depend on either one. The relevant distinction is among individual inference, task-bounded agentic trajectories and persistent agentic systems whose consequential state survives those trajectories, rather than between evolutionary and gradient-based learning. AI systems are already being pushed toward longer-lived state, less frequent human intervention, recovery from failure and adaptation to changing environments. Those properties are useful enough that persistent operation is likely to remain an important engineering direction even if the specific artificial-id architecture proposed here is not adopted. The evidence labels in Table 1 remain deliberately limited: two rows are Demonstrated, one is Implied and six are Not tested.
7
Conclusion
Differential persistence acting on controllers that received no task-specific behavioral objective produced adaptive sensorimotor control, selected an unintended physical strategy when that behavior prolonged persistence, and replaced a learned mapping during continued operation after its environmental meaning changed. Reassigning which environmental source was sustaining also redirected behavior without supplying a new behavioral objective. These results support a minimal experimental claim: persistence-organized adaptive control can emerge without a task-specific behavioral objective delivered to the controller. The architectural proposal separates such drive from general-purpose reasoning and couples them within a continuing agentic system. At that systems level, consequential state and drive that persist across task boundaries require an alignment boundary that persists with them, comprising trusted observations, consequence channels, persistent state, authority, identity, provenance and hard constraints. Whether this architecture scales remains open. A transferable drive would need to generalize across heterogeneous agentic environments while its system-level alignment boundary remains effective during persistent operation and continued adaptation.
AI Assistance Disclosure Generative AI tools were used to assist with writing and debugging experimental code, editing the manuscript, and generating illustrative figures. AI-generated text and code were reviewed and revised by the author before inclusion. The scientific arguments, experimental design, analysis, interpretation and conclusions are the author’s responsibility.
14
Artificial Id
A P REPRINT
References [1] [2] [3]
[4] [5]
[6]
[7] [8] [9]
[10]
[11] [12] [13] [14] [15] [16]
[17] [18] [19]
[20]
[21]
openclaw/openclaw. Sept. 2026. URL: https://github.com/openclaw/openclaw (visited on 09/09/2026). OpenHarness. 2026. URL: https://github.com/thu-nmrc/OpenHarness (visited on 09/10/2026). J. Zhang, S. Hu, C. Lu, R. Lange, and J. Clune. “Darwin Gödel machine: open-ended evolution of self-improving agents”. In: International Conference on Learning Representations. Vol. 2026. 2026, pp. 104223–104294. URL: https : / / proceedings . iclr . cc / paper _ files / paper / 2026 / hash / aa5f5e6eb6f613ec412f1d948dfa21a5-Abstract-Conference.html. X. Yin, X. Wang, L. Pan, L. Lin, X. Wan, and W. Y. Wang. Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement. arXiv:2410.04444 [cs.AI]. May 2025. DOI: 10.48550/arXiv.2410.04444. C. Fernando, D. Banarse, H. Michalewski, S. Osindero, and T. Rocktäschel. Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution. arXiv:2309.16797 [cs.CL]. Sept. 2023. DOI: 10.48550/arXiv.2309. 16797. L. A. Agrawal, S. Tan, D. Soylu, N. Ziems, R. Khare, K. Opsahl-Ong, A. Singhvi, H. Shandilya, M. J. Ryan, M. Jiang, C. Potts, K. Sen, A. Dimakis, I. Stoica, D. Klein, M. Zaharia, and O. Khattab. “Gepa: Reflective prompt evolution can outperform reinforcement learning”. In: International Conference on Learning Representations. Vol. 2026. 2026, pp. 8479–8565. URL: https://proceedings.iclr.cc/paper_files/paper/2026/ hash/0e9e708b6f48e14fd0ac29e167413f76-Abstract-Conference.html. T. Nakagaki, H. Yamada, and Á. Tóth. “Maze-solving by an amoeboid organism”. In: Nature 407.6803 (Sept. 2000), pp. 470–470. DOI: 10.1038/35035159. T. Saigusa, A. Tero, T. Nakagaki, and Y. Kuramoto. “Amoebae Anticipate Periodic Events”. In: Physical Review Letters 100.1 (Jan. 2008), p. 018101. DOI: 10.1103/PhysRevLett.100.018101. R. P. Boisseau, D. Vogel, and A. Dussutour. “Habituation in non-neural organisms: evidence from slime moulds”. In: Proceedings of the Royal Society B: Biological Sciences 283.1829 (2016), p. 20160446. URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC4855389/. C. R. Reid, T. Latty, A. Dussutour, and M. Beekman. “Slime mold uses an externalized spatial “memory” to navigate in complex environments”. In: Proceedings of the National Academy of Sciences 109.43 (Oct. 2012), pp. 17490–17494. DOI: 10.1073/pnas.1215037109. C. R. Reid. “Thoughts from the forest floor: a review of cognition in the slime mould Physarum polycephalum”. In: Animal Cognition 26.6 (Nov. 2023), pp. 1783–1797. DOI: 10.1007/s10071-023-01782-1. J. DiFrisco and R. Gawne. “Biological agency: a concept without a research program”. In: Journal of Evolutionary Biology 38.2 (2025), pp. 143–156. DOI: 10.1093/jeb/voae153. Y. P. Shkolnikov. From Deceptive Outputs to Deceptive Mechanisms: A Causal Framework for Language-Model Deception Research. arXiv:2609.04166 [cs.AI]. Sept. 2026. DOI: 10.48550/arXiv.2609.04166. langchain-ai/langgraph. original-date: 2023-08-09T18:33:12Z. Sept. 2026. URL: https://github.com/ langchain-ai/langgraph (visited on 09/10/2026). Helix: A Vision-Language-Action Model for Generalist Humanoid Control. Feb. 2025. URL: https://www. figure.ai/news/helix (visited on 09/09/2026). J. Geiping, S. McLeish, N. Jain, J. Kirchenbauer, S. Singh, B. Bartoldson, B. Kailkhura, A. Bhatele, and T. Goldstein. “Scaling up test-time compute with latent reasoning: A recurrent depth approach”. In: Advances in Neural Information Processing Systems 38 (2025), pp. 41340–41391. URL: https : / / proceedings . neurips . cc / paper _ files / paper / 2025 / hash / 3b01972cf31e6fa0fe29e4b8b5c2a0a1 - Abstract Conference.html. Safety and alignment in an era of long-horizon models. July 2026. URL: https://openai.com/index/ safety-alignment-long-horizon-models/ (visited on 09/10/2026). R. Shah and F. Flynn. Securing the future of AI agents. June 2026. URL: https://deepmind.google/blog/ securing-the-future-of-ai-agents/ (visited on 09/10/2026). P.-P. Grassé. “La reconstruction du nid et les coordinations interindividuelles chez Bellicositermes natalensis et Cubitermes sp. la théorie de la stigmergie: Essai d’interprétation du comportement des termites constructeurs”. In: Insectes Sociaux 6.1 (Mar. 1959), pp. 41–80. DOI: 10.1007/BF02223791. F. Heylighen. “Stigmergy as a universal coordination mechanism I: Definition and components”. In: Cognitive Systems Research 38 (2016), pp. 4–13. URL: https://www.sciencedirect.com/science/article/pii/ S1389041715000327. T. Johannink, S. Bahl, A. Nair, J. Luo, A. Kumar, M. Loskyll, J. A. Ojea, E. Solowjow, and S. Levine. “Residual reinforcement learning for robot control”. In: 2019 international conference on robotics and automation (ICRA). IEEE, 2019, pp. 6023–6029. URL: https://ieeexplore.ieee.org/abstract/document/8794127/. 15
Artificial Id
A P REPRINT
[22] N. Bredeche, E. Haasdijk, and A. Prieto. “Embodied evolution in collective robotics: a review”. In: Frontiers in Robotics and AI 5 (2018), p. 12. URL: https://www.frontiersin.org/journals/robotics- andai/articles/10.3389/frobt.2018.00012/full. [23] A. Steyven, E. Hart, and B. Paechter. “Understanding Environmental Influence in an Open-Ended Evolutionary Algorithm”. In: Parallel Problem Solving from Nature – PPSN XIV. Ed. by J. Handl, E. Hart, P. R. Lewis, M. López-Ibáñez, G. Ochoa, and B. Paechter. Vol. 9921. Cham: Springer International Publishing, 2016, pp. 921– 931. DOI: 10.1007/978-3-319-45823-6_86. [24] L. Floridi. “AI as Agency without Intelligence: On Artificial Intelligence as a New Form of Artificial Agency and the Multiple Realisability of Agency Thesis”. In: Philosophy & Technology 38.1 (Mar. 2025), p. 30. DOI: 10.1007/s13347-025-00858-9. [25] Z. Kenton, R. Kumar, S. Farquhar, J. Richens, M. MacDermott, and T. Everitt. “Discovering agents”. In: Artificial Intelligence 322 (2023), p. 103963. URL: https://www.sciencedirect.com/science/article/pii/ S0004370223001091. [26] G. Seifert, A. Sealander, S. Marzen, and M. Levin. “From reinforcement learning to agency: Frameworks for understanding basal cognition”. In: BioSystems 235 (2024), p. 105107. DOI: 10.1016/j.biosystems.2023. 105107. [27] D. Floreano, P. Husbands, and S. Nolfi. “Evolutionary Robotics”. In: Springer Handbook of Robotics. Springer Verlag, 2008. URL: https : / / infoscience . epfl . ch / bitstreams / 3cbfa404 - 5de4 - 442d - bf98 1963dcd2e011/download. [28] S. Singh, R. L. Lewis, A. G. Barto, and J. Sorg. “Intrinsically motivated reinforcement learning: An evolutionary perspective”. In: IEEE transactions on autonomous mental development 2.2 (2010), pp. 70–82. URL: https: //ieeexplore.ieee.org/abstract/document/5471106/. [29] M. Keramati and B. Gutkin. “Homeostatic reinforcement learning for integrating reward collection and physiological stability”. In: Elife 3 (2014), e04811. DOI: 10.7554/eLife.04811. [30] A. S. Rao and M. P. Georgeff. “BDI agents: From theory to practice”. In: ICMAS. 1995, pp. 312–319. URL: https://cdn.aaai.org/ICMAS/1995/ICMAS95-042.pdf. [31] P. Jiang, J. Lin, Z. Shi, Z. Wang, L. He, Y. Wu, M. Zhong, P. Song, Q. Zhang, H. Wang, X. Xu, H. Xu, P. Han, D. Zhang, J. Sun, C. Yang, K. Qian, T. Wang, C. Hu, M. Li, et al. Adaptation of Agentic AI: A Survey of Post-Training, Memory, and Skills. arXiv:2512.16301 [cs.AI]. Mar. 2026. DOI: 10.48550/arXiv.2512.16301. [32] K. Khetarpal, M. Riemer, I. Rish, and D. Precup. “Towards continual reinforcement learning: A review and perspectives”. In: Journal of Artificial Intelligence Research 75 (2022), pp. 1401–1476. URL: https://www. jair.org/index.php/jair/article/view/13673. [33] S. Dohare, J. F. Hernandez-Garcia, Q. Lan, P. Rahman, A. R. Mahmood, and R. S. Sutton. “Loss of plasticity in deep continual learning”. In: Nature 632.8026 (2024), pp. 768–774. URL: https://www.nature.com/ articles/s41586-024-07711-7. [34] J. Bauer, K. Baumli, F. Behbahani, A. Bhoopchand, N. Bradley-Schmieg, M. Chang, N. Clay, A. Collister, V. Dasagi, L. Gonzalez, et al. “Human-timescale adaptation in an open-ended task space”. In: International Conference on Machine Learning. PMLR, 2023, pp. 1887–1935. URL: https://proceedings.mlr.press/ v202/bauer23a.html. [35] Z. Dulberg, R. Dubey, I. M. Berwian, and J. D. Cohen. “Having multiple selves helps learning agents explore and adapt in complex changing worlds”. In: Proceedings of the National Academy of Sciences 120.28 (July 2023), e2221180120. DOI: 10.1073/pnas.2221180120. [36] D. Ackley and M. Littman. “Interactions between learning and evolution”. In: Proceedings of the Second Conference on Artificial Life, 1991. Addison-Wesley, 1991. URL: https : / / cir . nii . ac . jp / crid / 1573950400004246144. [37] A. Gupta, S. Savarese, S. Ganguli, and L. Fei-Fei. “Embodied intelligence via learning and evolution”. In: Nature communications 12.1 (2021), p. 5721. URL: https://www.nature.com/articles/s41467-021-25874-z. [38] A. Lynch, J. Hughes, A. Serrano, R. Kirk, and S. Bowman. Agentic Misalignment in Summer 2026. July 2026. URL : https://alignment.anthropic.com/2026/agentic-misalignment-summer-2026/ (visited on 09/10/2026). [39] T. Everitt, M. Hutter, R. Kumar, and V. Krakovna. “Reward tampering problems and solutions in reinforcement learning: a causal influence diagram perspective”. In: Synthese 198.S27 (Nov. 2021), pp. 6435–6467. DOI: 10.1007/s11229-021-03141-4. [40] S. M. Omohundro. “The Basic AI Drives”. In: Proceedings of the 2008 conference on Artificial General Intelligence. NLD: IOS Press, June 2008, pp. 483–492.
16
Artificial Id
A P REPRINT
[41] D. Hendrycks. Natural Selection Favors AIs over Humans. arXiv:2303.16200 [cs.CY]. July 2023. DOI: 10. 48550/arXiv.2303.16200. [42] S. Cohen, R. Bitton, and B. Nassi. Here Comes The AI Worm: Unleashing Zero-click Worms that Target GenAI-Powered Applications. arXiv:2403.02817 [cs.CR]. Jan. 2025. DOI: 10.48550/arXiv.2403.02817. [43] M. Kinniment, L. J. K. Sato, H. Du, B. Goodrich, M. Hasin, L. Chan, L. H. Miles, T. R. Lin, H. Wijk, J. Burget, A. Ho, E. Barnes, and P. Christiano. Evaluating Language-Model Agents on Realistic Autonomous Tasks. arXiv:2312.11671 [cs.CL]. Jan. 2024. DOI: 10.48550/arXiv.2312.11671.
17