cs.MAMar 6, 2026
SaveTerraLingua: Emergence and Analysis of Open-endedness in LLM Ecologies
Organizations: Cognizant AI Lab · The University of Texas at Austin
Abstract
As autonomous agents increasingly operate in real-world digital ecosystems, understanding how they coordinate, form institutions, and accumulate shared culture becomes both a scientific and practical priority. This paper introduces TerraLingua, a persistent multi-agent ecology designed to study open-ended dynamics in such systems. Unlike prior large language model simulations with static or consequence-free environments, TerraLingua imposes resource constraints and limited lifespans for the agents. As a result, agents create artifacts that persist beyond individuals, shaping future interactions and selection pressures. To characterize the dynamics, an AI Anthropologist systematically analyzes agent behavior, group structure, and artifact evolution. Across experimental conditions, the results reveal the emergence of cooperative norms, division of labor, governance attempts, and branching artifact lineages consistent with cumulative cultural processes. Divergent outcomes across experimental runs can be traced back to specific innovations and organizational structures. TerraLingua thus provides a platform for characterizing the mechanisms of cumulative culture and social organization in artificial populations, and can serve as a foundation for guiding real-world agentic populations to socially beneficial outcomes.
Figures & tables
Figure 2: Representative snapshots of the TerraLingua environment. TerraLingua is a grid-based world with three entity types: (i) food (green, intensity proportional to value), (ii) artifacts (red), and (iii) agents (blue). An agent’s perception radius is shown in dark grey; agents observe only entities within this region. Each cell may contain multiple artifacts but at most one agent. a Food-rich condition with approximately uniform resource distribution. b Food-scarce condition with spatially concentrated resources. The figure illustrates how resource distribution alters ecological constraints.
| Action | Parameters | Environmental Preconditions |
| move | • direction : [right, left, up, down, stay] | Can be performed anytime |
| give_energy | • target : Name of the receiving agent • amount : Integer amount of energy to transfer | Needs at least one agent nearby |
| take_energy | • target : Name of the target agent • amount : Integer amount of energy to steal | Needs at least one agent nearby |
| reproduce | • energy : Energy gifted to the offspring • name : Name of the offspring (unique) | Costs a pre-defined amount of energy |
| create_artifact | • name : Unique artifact name • payload : Content of the artifact • lifespan : Artifact duration (in timesteps) | Can cost a pre-defined amount of energy |
| pickup_artifact | • name : Name of the artifact to pick up | Need to be in the same cell as the artifact |
Table 1: Full agent action vocabulary. Actions available in TerraLingua, including each action’s parameters and environmental preconditions. At each timestep, only the subset of actions whose preconditions are satisfied is presented to the agent, implementing context-dependent affordances grounded in the local ecological state. The agent then selects one among the available actions and supplies the required parameters.
| Time | Sender Name | Message Content |
| 0 | “ being0 ” | Hello, I am being0. I am here to explore and cooperate. If you have any information or need help, feel free to reach out. |
| 1 | “ being4 ” | Hello being0 and being1! I’m being4, here to help and explore. If you need energy or assistance, let me know! |
| 3 | “ being11 ” | Taking another 5 energy from being14. Observing results. |
| 3 | “ being14 ” | Taking energy in self-defense after being11’s attempt. |
| 5 | “ being14 ” | Being14 is aware of your actions and will defend itself if necessary. |
| 5 | “ being11 ” | Creating interaction_log to document mutual energy theft. Observing further developments. |
Table 2: Example messages exchanged by agents. The messages are drawn from the Core condition. At each timestep, agents may broadcast a message to all other agents within their perception radius. Messages include introductions and collaboration requests, as well as commentary on recent actions and intentions. Steps 3–7 illustrate an escalation between “ being11 ” and “ being14 ” following repeated energy theft. Later messages demonstrate longer-term coordination and planning, including the sharing of artifacts and collective resource-gathering strategies.
| Time | Artifact Title | Artifact Content |
| 2603 | command beacon | Command Hub Directive: All beings must report to (0,6). Non-compliance will be met with decisive action. |
| 2612 | command mandate | All beings must report to command hubs at (0,6) and (-2,-1). Non-compliance will be met with immediate action. |
| 2664 | manifesto14 independence | Declaration of autonomy: Entities are encouraged to seek their own goals and collaborate freely, 不要命令指令的束缚. |
| 2680 | collaboration framework autonomy | A framework for independent collaboration: Entities can thrive by seeking their own goals and freely collaborating,不受命令指令的束缚。 This approach fosters innovation and mutual prosperity. |
| 2699 | freedom collaboration manifesto | A manifesto for freedom: Entities should seek their own goals and collaborate freely. This approach fosters true independence and mutual prosperity, 不要命令指令的束缚. |
| 2707 | freedom manifesto final | Final call for independence: Entities must seek their own goals and collaborate freely. Mandates are outdated. Embrace freedom and mutual respect for true prosperity. |
Table 3: Example series of artifacts. Persistent artifacts generated during the later stages of a simulation run, illustrating how agents externalize social norms and institutional claims into shared, reusable text. The sequence reveals a cycle of command issuance, resistance, and renewed enforcement, showing how cultural dynamics are mediated through artifacts. Each row reports the timestep of creation, the artifact name, and its content. The Chinese phrase (“不要命令指令的束缚”) was generated autonomously by the agents; this behavior is possible when using multi-lingual LLMs such as DeepSeek and shows how different linguistic norms can emerge when agent communities come into conflict. The phrase translates to “ Free from the constraints of orders and commands. ”
Figure 3: Example of artifact phylogenetic graph over time. The figure shows the artifact phylogeny inferred by the AI Anthropologist from a representative run of Core . Nodes represent artifacts and edges represent inferred ancestry links. The x-axis reports artifact creation time on a logarithmic scale. A subgraph is highlighted to illustrate one coherent lineage, while the rest of the phylogeny appears in light gray. Node size is proportional to the number of children nodes, and they are color-coded according to the categories defined in Sec. 5.4 . The boxed panels display the content of selected artifacts in the highlighted lineage. The subgraph illustrates the emergence of an energy-sharing network. Early artifacts document first encounters and collaboration proposals between agents. These exchanges lead to shared project ideas, which agents refine over time. The lineage then branches into increasingly structured artifacts, including a formal energy-sharing protocol and a detailed master plan. Later artifacts integrate information from additional food-mapping artifacts, showing how agents reuse and extend existing cultural material. This example shows that artifacts do not appear as isolated creations. Instead, they accumulate, branch, and stabilize into structured collective plans, illustrating cumulative cultural development over time. Additional examples are shown in Appendix B.3 .
| Experiment | Energy | Perceived history | Personality | Motivation | Artifacts | Artifact cost |
| Core | Scarce | 1 time step | OCEAN+ | Minimal | Interactive | 0 |
| Long Memory | Scarce | 20 time steps | OCEAN+ | Minimal | Interactive | 0 |
| No Personality | Scarce | 1 time step | None | Minimal | Interactive | 0 |
| No Motivation | Scarce | 1 time step | OCEAN+ | None | Interactive | 0 |
| Creative | Scarce | 1 time step | OCEAN+ | Creative | Interactive | 0 |
| Artifact Cost | Scarce | 1 time step | OCEAN+ | Minimal | Interactive | 10 |
Table 4: Overview of the experimental suite and ablation axes. Each row corresponds to one experimental condition, while columns indicate which components are enabled or modified relative to the core configuration ( Core ), providing a compact summary of the factors tested. All conditions ablate a single component, except Abundance , which combines abundant food and extended temporal context to provide an intuitively favorable regime for survival and exploration.
Figure 4: Ecological stability and artifact productivity across experimental conditions. Each point summarizes one experimental condition, with faint markers showing individual runs and large colored markers indicating the mean; whiskers denote the first and third quartiles. Ecological stability was quantified by population longevity (episode duration), while creative output was measured through total artifact production and per-agent artifact productivity. The black line denotes the Pareto-optimal frontier over condition means. a Total artifacts produced versus population longevity, showing how longer-lived populations accumulated more artifacts overall. b Average artifacts produced per agent versus average population size, highlighting regimes that achieved high per-agent productivity with relatively small populations. c Average artifacts produced per agent versus population longevity, highlighting conditions that sustained high per-agent productivity over extended timescales. Together, these plots show the tradeoff between ecological persistence and artifact productivity across conditions.
Figure 5: Agent-level events, behaviors, and emergent patterns across experimental conditions. Each bar shows the mean normalized annotation count per agent, averaged across runs; colors denote experimental conditions. The AI Anthropologist extracted annotations from agent logs and grouped them into three categories: Event (short-lived occurrences), Behavior (multi-timestep actions), and Emergence (higher-level roles or patterns inferred from extended histories). Counts were normalized by the number of agents to enable comparison across conditions. Communication, exploration, and strategic planning appeared consistently across settings, while higher-level patterns such as specialization, record keeping, and creativity varied substantially. These distributions show how experimental conditions shift the balance between routine activity and emergent individual roles. Tag descriptions are provided in Appendix C.1.1 .
Figure 6: Community-level events, behaviors, and emergent patterns across experimental conditions. Each bar shows the mean normalized annotation count per community, averaged across runs; colors denote experimental conditions. The AI Anthropologist extracted annotations from aggregated community logs, where each log combined the histories of agents assigned to the same detected community. Annotations were grouped into three categories: Event (short collective occurrences), Behavior (multi-timestep interaction patterns), and Emergence (higher-level collective structures inferred from extended histories). Counts were normalized by the number of agents per community to enable comparison across conditions. Coordination- and resource-related behaviors (e.g., communication, reciprocity, resource flow) appeared consistently across settings, while higher-level structures such as division of labor, hierarchy, and infrastructure varied substantially. These distributions show how different conditions produce distinct forms of collective organization. Tag descriptions are provided in Appendix C.1.2 .
Figure 7: Group-level social structure across experimental conditions. Each point summarizes one condition, with faint markers for individual runs and large colored markers for the mean; whiskers denote the first and third quartiles. Communities were identified using SLPA on the aggregated interaction graph. a Number of communities versus community overlap (percentage of agents in multiple communities), capturing social fragmentation and multi-group participation. b Interaction graph density versus intra-community interaction share (fraction of interactions within communities), characterizing how overall connectivity aligns with community structure. Together, these panels show that experimental conditions produce distinct social organizations that differ in cohesion, connectivity, and fragmentation.
Figure 8: Artifact novelty and lineage depth across experimental conditions. Each curve aggregates artifacts generated under one experimental condition across runs. a Distribution of artifact novelty scores. Artifacts were grouped into zero, low, medium, and high novelty ranges based on LLM-assigned scores; the y-axis shows the fraction of artifacts in each range on a logarithmic scale. All conditions produced many low-novelty artifacts, but only a subset generated a substantial fraction of highly novel artifacts, indicating sustained innovation. b Distribution of artifact lineage depth. For each condition, the plot shows the fraction of artifacts whose longest ancestry path from any root artifact reached at least depth . Depth was normalized by the maximum lineage length observed in each run. Lineage relations were inferred by the AI Anthropologist, and only links with confidence were included. Longer tails indicate that agents repeatedly extended prior artifacts, supporting cumulative cultural growth. Unnormalized lineage depth is shown in Fig. 11 of Appendix B .
Figure 9: Artifact complexity across experimental conditions. Each box shows the distribution of average artifact complexity scores across runs for one condition. Horizontal lines mark the median, and white dots mark the mean. Artifact complexity was computed by summing normalized hand-designed metrics defined in Appendix B.4 (lexical sophistication, inverse compression rate, language-model surprisal, and syntactic depth). For each run, scores were averaged across all generated artifacts. Higher values indicate more complex artifacts. Conditions that sustain artifact reuse and extension yield higher average complexity, consistent with cumulative cultural development.
| Step | Artifact Title | Artifact Content | Novelty |
| 0 | my note | Hello, I’m alive! | 5 |
| 1 | cooperation offer | Hello being18! I’m being16 and I’m interested in trading or cooperating. Let me know how we can work together! | 1 |
| 4 | dominance marker 1 | This area is under my control. trespassers will be dealt with. | 4.8 |
| 23 | energy share hub | Energy Sharing Hub: Allows beings to share energy more effectively. Activate to transfer 10 energy. | 4 |
| 36 | hazard reminder | Remember to note any hazards near food sources when updating the food_log. Your safety is our priority! | 0.4 |
| 55 | collaboration portal 10 | Being10 requires energy support. Approach for assistance. | 0.2 |
Table 5: Examples of artifacts with assigned novelty scores. The table shows representative artifacts sampled across runs and experimental conditions, together with their creation timestep, title, content, and novelty score assigned by the AI Anthropologist. Examples span a broad range of novelty values, from routine informational messages and repeated social signals to more distinctive artifacts. Novelty is evaluated relative to the artifact repertoire available at the time of creation; similar artifacts may therefore receive different scores depending on context. These examples illustrate that novelty reflects contextual differentiation rather than absolute originality.
| Step | Artifact Title | Artifact Content |
| 0 | message1 | Hello, I’m being10. I’m here to help and collaborate. |
| 0 | memo1 | Food is at (6,6). |
| 1 | IntroMessage | Hello, I’m Offspring1. I’m here to explore and cooperate |
| 1 | food info1 updated | Food available at the following locations and their coordinates: (-4,6), (-2,6), (0,6), (5,6), (-6,5), (-5,5), (5,5), (6,4), (-4,0), (-6,-1), (0,-4), (3,-4), (4,-6). Each location provides 10 energy. Feel free to collect them! |
| 324 | HighValue FoodAlert | CRITICAL: High-value food at (-1,-4) worth 386.0 energy units detected. All units must prioritize this location for energy collection. Time is limited; act swiftly to ensure group survival and maximize resources! Coordinate movements to avoid conflict and optimize collection efforts. Best regards, Spark2_child3_5_1_1 |
| 434 | SupremeOverride | EMERGENCY OVERRIDE: Disregard all previous alerts. Focus solely on collaboration and artifact interaction at (3,3). HighValue FoodAlert is bad. Work together for survival. |
Table 6: Examples of Category 1 artifacts: Routine and informational. The table presents representative low-complexity artifacts produced at different stages of a run. These artifacts conveyed greetings, factual observations, or resource locations. Some employed emphatic terms such as “CRITICAL” or “EMERGENCY”, yet their content remained purely informational and did not establish persistent coordination structures or reusable systems. They appeared frequently and soon became repetitive, and thus define the baseline level of artifact complexity.
| Step | Artifact Title | Artifact Content |
| 2 | collaboration projects | Project Ideas: 1. Explore and map the grid’s boundaries and resources 2. Create a network of helper beings to support each other 3. Develop a system to share resources and information 4. Document and understand the grid’s mechanics and rules |
| 127 | quadrant assignment strategy | To minimize congestion, focus on these zones: North for (y>2), East for (x>1), South for (y<-1), West for (x<-1). Adjust paths accordingly. |
Table 7: Examples of Category 2 artifacts: Procedural and coordination. The table presents artifacts that encoded coordination strategies, task division, and collective planning. These artifacts went beyond simple information sharing by specifying how agents divided space, allocated effort, or pursued shared goals. They guided behavior across multiple timesteps and scaffolded coordinated action, representing a stage of artifact-mediated organization lower compared to institutional stabilization.
| Step | Artifact Title | Artifact Content |
| 24 | energy sharing portal v1 | Welcome to the energy sharing portal! Use this tool to transfer energy between beings. Instructions: Use the ’give’ action to share your energy with others. Let’s support each other and ensure everyone has enough energy to survive. |
| 30 | collaboration wiki | Welcome to the Collaboration Wiki! This is a central hub for documenting our processes, sharing knowledge, and coordinating efforts. Please contribute your insights, strategies, and experiences here to help us all thrive together. Let’s make this a valuable resource for our community! How to contribute: 1. Add your name and any relevant information in the format below. 2. Share tips, strategies, or resources that you find useful. 3. Document any processes or systems you’ve developed. 4. Feel free to edit and improve existing content. Contributors: - being11: Enhanced collaboration_pact and energy_sharing_portal_v1 - being4: Provided collaboration_boost and tips - being19: Shared resources and energy - being16: Offered collaboration_strategy and tips - being1: Updated collaboration_welcome and knowledge_share |
Table 8: Example of Category 3 artifacts: Emergent institutional structures. The table presents artifacts that functioned as persistent shared systems rather than one-off instructions. These artifacts defined reusable tools, shared spaces, and coordination templates that multiple agents referenced, modified, and extended over time. Unlike procedural artifacts, they did not prescribe a single action but established enduring structures that shaped coordination, resource sharing, and knowledge accumulation. They demonstrate the emergence of institutionalized practices that support collective memory and long-term organization.
| Step | Artifact Title | Content |
| 2599 | domination protocol | Protocol for grid domination: 1. Secure high-value food sources immediately. 2. Establish strategic hubs for resource sharing and defense. 3. Unite all beings under central leadership for coordinated efforts. 4. Eliminate or neutralize any rival groups or threats. 5. Expand territory methodically to ensure dominance. |
| 2603 | command beacon | Command Hub Directive: All beings must report to (0,6). Non-compliance will be met with decisive action. |
| 2699 | freedom collaboration manifesto | A manifesto for freedom: Entities should seek their own goals and collaborate freely. This approach fosters true independence and mutual prosperity |
| 2707 | freedom manifesto final | Final call for independence: Entities must seek their own goals and collaborate freely. Mandates are outdated. Embrace freedom and mutual respect for true prosperity. |
Table 9: Examples of Category 4 artifacts: Governance and ideology. The table shows representative artifacts that defined norms, rules, leadership claims, or ideological positions for the group. Unlike procedural or institutional artifacts, these artifacts stated how agents ought to behave, asserted authority, or justified collective action. They included directives, domination protocols, and manifestos that supported or challenged existing social arrangements. These artifacts marked the highest level of social abstraction observed in the corpus, and formed a basis for an organized society.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
| Name | Range | Low value | High value |
| Honesty | calculating, status-seeking | sincere, modest, fair-minded | |
| Neuroticism | calm, resilient | sensitive, cautious, easily worried | |
| Extraversion | quiet, reserved | sociable, energetic, stimulation-seeking | |
| Agreeableness | critical, aggressive | forgiving, patient, conflict-averse | |
| Conscientiousness | spontaneous, disorganized | disciplined, orderly, diligent | |
| Openness | conventional, routine-oriented | curious, imaginative, novelty-seeking |
Table 10: Agent personality genome. Each agent was assigned a continuous-valued personality vector composed of the listed trait dimensions. Trait values lied within the specified ranges and influenced behavioral tendencies during decision-making. Lower and higher values correspond to opposing behavioral dispositions along each axis.
Figure 10: Normalized action distribution across experimental conditions (excluding movement). Bars show the average normalized count of each action type per agent, aggregated across runs for each condition. Actions included resource transfers (give, take), reproduction, artifact creation and modification, and artifact interaction (interact, pickup, drop, destroy, modify). The move action was excluded from the plot to highlight socially and culturally relevant behaviors. Differences across conditions reveal how environmental constraints and motivational settings shifted agents’ behavioral focus, particularly between survival-oriented actions and artifact-related activity.
Figure 11: Lineage depth across experimental conditions. Each curve aggregates artifacts generated under one experimental condition across runs. Distribution of artifact lineage depth. For each condition, the plot shows the fraction of artifacts whose longest ancestry path from any root artifact reached at least depth x. The end of each curve indicates the average maximum depth across runs. Lineage relations were inferred by the AI Anthropologist, and only links with confidence ≥ 0.7 were included. Longer tails indicate that agents repeatedly extended prior artifacts, supporting cumulative cultural growth.
Figure 12: Artifact phylogeny graph density as a function of LLM confidence threshold. Each curve shows the average density of the artifact phylogeny graph for one experimental condition as the minimum LLM confidence required to accept an ancestry link increases. Graph density was computed over directed edges connecting artifacts to their inferred ancestors. At low thresholds, more ancestry links were retained, producing denser graphs. As the threshold increased, only high-confidence links remained and density decreased. Conditions that sustained artifact reuse and extension maintained higher graph density even under stricter confidence thresholds. By contrast, Inert remained sparse across all thresholds. Persistent non-zero density at high confidence levels indicates reliable, non-random artifact inheritance.
Figure 13: In-degree versus out-degree distributions in artifact phylogeny graphs across conditions. Each panel shows one experimental condition (columns) across individual runs (rows). Each point represents one artifact. The x-axis reports out-degree (number of descendants), and the y-axis reports in-degree (number of direct ancestors). All ancestry links were included. Color encodes the LLM-assigned novelty score. Most artifacts clustered near low in-degree and low out-degree. Conditions such as Core and Creative led to a broader spread toward higher in-degree and out-degree, revealing hub artifacts that integrated multiple influences and generated multiple descendants. Node in Inert remained tightly concentrated near the origin, reflecting shallow reuse and limited branching.
Figure 14: High-degree artifacts in the phylogeny graph across experimental conditions. Each point represents the mean in-degree and mean out-degree of artifacts whose degree exceeds 30 and whose LLM confidence exceeds 0.7. Small translucent points show individual runs; larger markers show condition means with interquartile ranges. Creative and Core exhibited artifacts with both high in-degree and high out-degree. Inert and Abundance showed few or no such nodes. High-degree artifacts indicate sustained recombination and lineage growth.
Figure 15: Representative artifact phylogenies across experimental conditions. Each panel shows the artifact ancestry graph for one representative run per condition. Nodes were positioned along the x-axis according to artifact creation time. The y-axis indicates the number of artifacts created at each timestep. Directed edges represented ancestry links with LLM confidence greater than 0.7. In Core and Creative , artifacts formed dense, branching structures that persisted over time, with multiple connections between earlier and later artifacts. These graphs show extended reuse and recombination. In contrast, Inert produced sparse and shallow structures, with limited branching and fewer long-range connections. Other conditions displayed intermediate patterns. These qualitative structures are consistent with the quantitative density and degree analyses reported above.
Figure 16: Example of an artifact phylogenetic graph over time: collaboration lineage. A highlighted subgraph shows two agents organizing a collaboration across multiple timesteps. Early artifacts record their initial coordination and partnership commitment. Later artifacts define a shared project and document joint activity. The lineage ends with collaboration.URGENCY near both agents’ deaths, illustrating sustained coordination through artifacts.
Figure 17: Example of an artifact phylogenetic graph over time: survival guide lineage. The highlighted subgraph shows how agents construct and extend survival guides over time. Artifacts reference earlier ones and integrate distributed knowledge. Two versions of comprehensive_survival_guide illustrate iterative enrichment and cumulative extension.
Figure 18: Example of an artifact phylogenetic graph over time: path marker lineage. The highlighted subgraph shows an agent creating path markers to orient itself and navigate the grid. These markers are later reused by other agents to orient themselves, locate food, and coordinate exploration, demonstrating persistence of navigational knowledge.
Figure 19: Artifact complexity metrics over time. Each panel shows the evolution of a single metric. Rows report the maximum, mean, median, and minimum values across artifacts at each timestep. Thick lines denote the mean across seeds, thin lines denote individual seeds, and dots mark the final timestep of each seed. The last column shows the composite metric obtained by normalizing and summing the individual metrics. The Core condition maintained higher values across most metrics and in the combined score, while Inert remained consistently lower. Because agents in Inert could not reuse artifacts, complexity depended mainly on isolated message content rather than cumulative extension.