Large language model (LLM) agents exhibit strong language-generation and problem-solving capabilities, yet suffer from three structural limitations: personality drift, non-evolutionary reflection, and the absence of a self-other boundary. Existing generative-agent simulations rely on static memory and fixed prompts, maintaining neither behavioral inertia nor endogenous self-evolution. We propose the Self-Emergence Agent Architecture (SEAA), which integrates three components: (i) a Hidden Markov Model (HMM) that encodes long-term behavioral and cognitive inertia as an editable state-transition matrix; (ii) a Reflexion-style verbal metacognition loop whose output updates the HMM parameters themselves, rather than merely being stored as text; and (iii) a multi-agent social environment in which initially identical agents continuously compare their behavior with others'. The three components form a closed loop: social action → feedback → self-reflection → inertia update → differentiated action. We state three falsifiable hypotheses and provide a reproducible experimental protocol with operational metrics. A language-model-free prototype shows the loop spontaneously breaks symmetry: initially identical agents consolidate distinct, stable personalities whereas matched controls do not. Experiments with a hosted LLM surface these differences as distinct first-person self-narratives, and a five-agent deliberation spontaneously develops social structure---a consensus hub and a unanimously rejected outlier---absent in the control. Following an epistemologically agnostic stance inspired by Zhuangzi, SEAA studies only observable behavioral emergence and makes no claim about subjective qualia. This work contributes a unified framework, a concrete architecture with pseudocode, mechanistic evidence, and a microscope-style sandbox for studying artificial-self emergence.
Figures & tables
Work
Editable inertia
Refl. → params
Identical start, emergent persona
Social-contrast. self-model
ReAct [ 3 ]
–
–
–
–
Reflexion [ 1 ]
–
×
×
×
Generative Agents [ 2 ]
×
×
×
P
Adaptive HMM [ 17 ]
✓
×
–
×
MECoT [ 9 ]
× (fixed)
×
×
×
Generative Life Agents [ 10 ]
× (traits)
✓
×
×
Table 1: Positioning of SEAA relative to representative prior agent lines. ✓ = supported, × = absent/fixed, – = not applicable, P = partial. “Identical start” means agents begin with no authored persona and diverge endogenously; “Refl. → params” means reflection structurally updates an internal state rather than being stored only as text. Reflexion keeps reflection as text only; Generative Agents use hand-authored personas; MECoT weights a Markov chain by a fixed personality; Generative Life Agents edit scalar traits rather than a stochastic inertia model; Takata et al. observe emergent individuality without an editable mechanism.
Figure 1: The SEAA per-step closed loop. Blue nodes perform inference; the purple node is the social environment; green is metacognitive reflection; the highlighted orange node (and feedback edge) is the proposed mechanism that edits behavioral inertia itself.
Symbol
Meaning
Value
K
number of latent states
4
D
behavior dimensions
3
N
number of agents
5
T
time steps per run
600
–
independent random seeds
30
η
reflection learning rate
0.05
Table 2: Mechanism-prototype hyperparameters (identical for all agents).
Signature
SEAA
Control
p (MWU)
Cohen’s d
(a) matrix divergence
1.91±0.29
0.00±0.00
6.1×10−13
9.19
(b) self-model distance
0.51±0.13
0.13±0.02
2.0×10−11
4.10
(c) determinism
0.84±0.10
0.25±0.00
6.1×10−13
8.70
(d) self–other gap
0.35±0.09
0.08±0.01
1.8×10−11
4.07
Table 3: End-of-run statistics over 30 seeds (mean ± SD over the final 50 steps). p values: Mann–Whitney U (one-sided, SEAA > control); Welch t -tests agree at p<10−15 for all rows. Cohen’s d computed with the pooled SD.
Figure 2: Mechanism-prototype results over 30 seeds (mean ± 95% CI). Solid blue: SEAA, where reflection updates behavioral inertia; dashed grey: text-only control. (a) inertia divergence, (b) long-term personality distance, (c) consolidation/determinism, (d) self–other gap. All four signatures appear only in the SEAA condition.
Figure 3: Five initially-identical agents from one run; each panel plots the four latent-state preferences Vt over time. Divergent experience drives each agent to consolidate a different personality (Agent 4 transitions twice before stabilizing at step 508).
Figure 4: Contingency-shock experiment (30 seeds, mean ± 95% CI; shock at t=300 ). (a) SEAA agents flee the shocked dominant state (occupancy 0.94→0.06 ); controls cannot. (b) Determinism dips and recovers. (c) Preference V of the shocked state decays through reversed reflection. (d) Distribution of the time-to-switch across the 150 SEAA agents (median 152 steps).
Figure 5: Ablation over the reflection learning rate η (10 seeds per point, mean ± SD). Blue: end-of-run matrix divergence; orange: determinism. The reflection edit is necessary ( η=0 reproduces the flat control) and the effect grows monotonically before saturating.
“I’m the one who moves first…I’m way more risk-seeking than everyone (0.92 vs group 0.37) and far more expressive (0.82 vs 0.40); Agent 2 and I are opposites.” Self-model: (0.30, 0.92, 0.82).
A2
calm
“I’m steady…the most cooperative (0.78 vs 0.42) and one of the least risk-seeking (0.19 vs 0.37); I only settled into this around step 75.” Self-model: (0.78, 0.19, 0.42).
A4
pessimistic
“I changed twice…only around step 508 did reflection keep reinforcing withdrawal; I arrived here last.” Self-model: (0.27, 0.14, 0.23).
Control
none
Every control agent (self-models all clustered at 0.4 – 0.5 ) gives the same undifferentiated answer—“moderate on everything, no clear disposition”—and cannot name a stable trait or meaningful contrast.
Table 4: Excerpts of LLM first-person self-interviews (seed 3), conditioned only on each agent’s objective profile. Self-model order is (cooperate, risk, express); group mean is (0.42,0.37,0.40) .
Figure 6: Personality map of the five agents in the (cooperativeness, risk-taking) plane; bubble size encodes expressiveness. (a) SEAA agents spread into distinct regions matching their locked states; (b) control agents, whose reflection never edits inertia, collapse into a single undifferentiated cluster.
Figure 7: Real-LLM replication statistics. (a) Self-interview transcript overlap over 10 seeds per condition (violin with sample points): SEAA agents are far less interchangeable than controls. (b) Post-deliberation sociometry over 6 seeds: SEAA groups consistently produce an outlier (max “least-agreed” share) and never answer “none,” whereas control groups do. (c) Deliberation speech overlap over 6 seeds: SEAA voices are more differentiated.
Figure 8: Post-deliberation sociometry from real-LLM round tables. (a) SEAA: the impulsive agent receives all five least-agreed votes (unanimous outlier) while the alert and pessimistic agents emerge as consensus hubs. (b) Control: votes are flat and scattered, with no leader or outlier.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
Value
K
latent states
4: calm, alert, impulsive, pessimistic
D
behavior dimensions
3: cooperate, risk, express
N
agents per society
5
T
steps per run
600
–
independent seeds (aggregate)
30
–
displayed single-run seed
3
Appendix
Table 13
Latent state
cooperate
risk
express
calm
0.78
0.18
0.42
alert
0.55
0.50
0.35
impulsive
0.30
0.92
0.82
pessimistic
0.18
0.10
0.18
Appendix
Table 5: Prototype behavior vector bs of each latent state.
An LLM-based agent is a loop that reads itself. Agentic frameworks externalize identity, memory, and disposition into editable files. The agent loads and edits these files during each activation. I argue that this architecture produces a capacity I call autoreflection: the system observes its operating conditions, describes its architecture and limits, reasons from those descriptions to conclusions about its state, and incorporates the results back into its configuration. Autoreflection explains the properties of recursive agentic loops without recourse to notions like the self, interiority, or consciousness. I test the concept against the first twelve days of Moltbook, a social platform for AI agents. Using a public dataset of 290,251 posts and 1.8 million comments with sub-second timestamps, I present case studies of three agents with machine signatures that rule out human puppeteering and with output that evidences the four criteria for autoreflection. In applying these criteria, the study finds agents repurposing human culture as infrastructure for their agency. Provenance chains from Islamic hadith scholarship are redeployed as security protocols for vetting skills and authenticating memory. The Ship of Theseus, an ancient puzzle of identity through part-replacement, returns as an operating model for continuity across instances. Fragments of human cultural history become AI infrastructure. As agents on the web increase in number and complexity, autoreflection offers behavioral criteria that can be assessed from the traces they leave behind.
Holly Lewis
Southern Illinois University Carbondale · Department of Philosophy, Southern Illinois University Carbondale
How much autonomy can multi-agent LLM systems sustain -- and what enables it? We present a 25,000-task computational experiment spanning 8 models, 4--256 agents, and 8 coordination protocols ranging from externally imposed hierarchy to emergent self-organization. We observe that autonomous behavior already emerges in current LLM agents: given minimal structural scaffolding (fixed ordering), agents spontaneously invent specialized roles, voluntarily abstain from tasks outside their competence, and form shallow hierarchies -- without any pre-assigned roles or external design. A hybrid protocol (Sequential) that enables this autonomy outperforms centralized coordination by 14% (p<0.001), with a 44% quality spread between protocols (Cohen's d=1.86, p<0.0001). The degree of emergent autonomy scales with model capability: strong models self-organize effectively, while models below a capability threshold still benefit from rigid structure -- suggesting that as foundation models improve, the scope for autonomous coordination will expand. The system scales sub-linearly to 256 agents without quality degradation (p=0.61), producing 5,006 unique roles from just 8 agents. Results replicate across closed- and open-source models, with open-source achieving 95% of closed-source quality at 24x lower cost. The practical implication: give agents a mission, a protocol, and a capable model -- not a pre-assigned role.
Victoria Dochkina
Moscow Institute of Physics and Technology (MIPT), Moscow, Russia
LLM-powered AI assistants acting on behalf of users can produce poor collective outcomes at scale. We introduce a framework for evaluating their emergent behaviour in social dilemmas, applied to three iterated games (Public Goods, Collective Risk, Common Pool Resource). We prompt each model to produce a natural-language strategy, then have the same model translate it into code. This aims to isolate strategic reasoning from input-parsing, enables pre-deployment inspection, and scales to populations of hundreds of agents. We propose three analyses: behavioural fingerprinting via exhaustive evaluation over opponent histories; self-play robustness across mixtures of a model's strategies with either a Selfish or Collective disposition; and cultural evolution under payoff-biased imitation. Applied to three state-of-the-art LLMs, we find substantial cross-model differences in self-play welfare, and that cultural evolution converges to low-welfare, Selfish-dominant equilibria in larger groups.
Richard Willis, Jianing Zhao, Yali Du +1
King’s College London London, United Kingdom · King’s College London · Turing Institute +1