TerraLingua: Agent Ecology & Culture

• TerraLingua is a multi-agent system where LLM-driven agents interact under resource constraints, enabling cumulative cultural evolution across generations.
• It employs persistent artifacts as external memory, allowing knowledge, norms, and governance proposals to shape ongoing social dynamics.
• The simulation balances ecological pressures and creative impulses to reveal how cooperation, division of labor, and institutional structures emerge.

TerraLingua is a persistent multi-agent ecology introduced to study open-ended dynamics in populations of LLM-driven agents operating under resource constraints and limited lifespans. In contrast to prior LLM simulations with static or consequence-free environments, TerraLingua makes artifacts persist beyond individuals, so that knowledge, norms, and tools can shape future interactions and selection pressures. Its reported outcomes include the emergence of cooperative norms, division of labor, governance attempts, and branching artifact lineages consistent with cumulative cultural processes, with divergent trajectories across runs traceable to specific innovations and organizational structures (Paolo et al., 6 Mar 2026).

1. Conceptual scope and research problem

TerraLingua is framed around a scientific and practical problem: as autonomous agents increasingly operate in real-world digital ecosystems, it becomes necessary to understand how they coordinate, form institutions, and accumulate shared culture. The system is designed as a controlled, interpretable, and experimentally tractable setting for studying open-ended dynamics, defined in the source as processes that produce continual novelty, increasing complexity, and cumulative culture (Paolo et al., 6 Mar 2026).

The platform is also a response to limitations identified in prior LLM-based agent simulations such as Interactive Simulacra and Sotopia. Four deficiencies are emphasized. First, static or consequence-free environments meant that behaviors had no lasting ecological cost. Second, the absence of resource limitations disconnected survival and reproduction from non-equilibrium ecological dynamics. Third, there was no true persistence or cumulative culture, because no external memory allowed conventions or tools to be stably externalized across generations. Fourth, diversity and creativity were often scaffolded through predefined objectives or role assignments rather than emerging from agent-environment interaction and endogenous motivation. TerraLingua addresses these gaps by combining durable artifacts, ecological constraints, and long-term population turnover (Paolo et al., 6 Mar 2026).

Within that framing, the system is not only a simulator of social interaction. It is intended as a platform for characterizing the mechanisms of cumulative culture and social organization in artificial populations, and as a possible foundation for guiding real-world agentic populations toward socially beneficial outcomes.

2. Ecological architecture

The TerraLingua environment is a 2D toroidal grid, a topology described as well studied in ALife, in which agents, food, and artifacts are spatially embodied. Food consists of discrete, spatially distributed, stochastically decaying resources. Artifacts are persistent, uniquely named text objects. Each agent perceives local surroundings within a fixed radius and can observe entities and artifacts in its neighborhood. Environmental conditions vary across experiments: resource distribution can be scarce and concentrated or abundant and uniform, food decays with probability $p$ at each timestep, only one agent may occupy a cell, and multiple artifacts may occupy the same cell (Paolo et al., 6 Mar 2026).

Agents begin in randomly assigned positions with energy $\epsilon$, fixed lifespan $\tau$, OCEAN+ personality traits, and empty inventories. Energy is required for survival, decreases each timestep, and is replenished by consuming food. Lifespan is a hard limit, also decremented each timestep; death occurs when either energy or lifespan reaches zero. The OCEAN+ specification includes Honesty, Neuroticism, Extraversion, Agreeableness, Conscientiousness, Openness, and Dominance, each on $[-1,1]$, together with Fertility on $[0.5,1]$. These traits are mutable and heritable via reproduction with mutation, thereby sustaining behavioral diversity over time (Paolo et al., 6 Mar 2026).

Action affordances are explicitly constrained by local context. Agents can move on the grid, give or take energy, reproduce, and create, modify, pick up, drop, give, or destroy artifacts. Energy exchange is limited to neighboring agents. Communication occurs by broadcast within the perception radius. In addition to public interaction, each agent maintains internal memory, described as a private field of persistent textual data supporting planning and self-reflection. Reproduction is asexual: the parent allocates energy to offspring and transmits a mutated personality genome (Paolo et al., 6 Mar 2026).

3. Artifacts, external memory, and open-endedness

A central design feature of TerraLingua is the use of persistent artifacts as external memory. Artifacts are named, text-based objects with editable content and user-defined duration or lifespan. They outlive individual agents and can be read, picked up, modified, destroyed, or transferred. They are visible to agents occupying or passing through the relevant cell. In functional terms, they store knowledge, norms, tools, and governance proposals (Paolo et al., 6 Mar 2026).

This persistence is tied directly to the platform’s account of cumulative cultural evolution. The paper identifies three mechanisms for open-endedness. Variation arises through spontaneous or context-sensitive innovation, personality effects, and artifact manipulation. Persistence occurs through artifact accumulation, whereby plans, information, and conventions remain in the environment. Cumulative change and lineage arise when agents modify or extend previous artifacts, generating branching, multi-generational cultural trajectories. The analytical target is therefore not novelty alone, but persistent branching artifact lineages, growing complexity, and functional or institutional differentiation (Paolo et al., 6 Mar 2026).

The ecological assumptions reinforce these mechanisms. Energy is scarce and consumed, so survival and reproduction are not guaranteed. That scarcity produces trade-offs among exploration, exploitation, and sociality. At the same time, agents’ internal parameters are fixed: all adaptation occurs through interaction, memory, and artifacts. The paper describes this as a decoupling of ontogeny and culture from phylogeny and genetics, aligning the design with dual-inheritance theory. It also treats artifact-mediated change as a form of niche construction, because artifacts alter the selective and interaction landscape for successors (Paolo et al., 6 Mar 2026).

4. The AI Anthropologist

Because open-ended simulations generate logs at a scale unsuitable for manual inspection, TerraLingua includes an analytical framework termed the AI Anthropologist. This component is an LLM-powered, non-intervening observer that analyzes behavior only after the simulation has completed, so that no feedback loop affects system dynamics (Paolo et al., 6 Mar 2026).

The framework operates at three levels. At the agent level, it tags and annotates individual behaviors and events. At the group level, it constructs social-interaction graphs whose edges encode encounter, communication, resource exchange, artifact exchange, and related relations in signed and weighted form. Communities are identified using SLPA, and emergent roles such as leadership and division of labor are inferred. At the artifact level, the system measures novelty and complexity, reconstructs phylogenetic relationships among artifacts, and classifies artifact roles ranging from routine notes to governance directives (Paolo et al., 6 Mar 2026).

Two quantitative constructions are especially important. Novelty is scored on a $[0,5]$ scale by an LLM, with multiple iterations averaged for robustness. Complexity is computed from lexical sophistication, compressed size, language-model surprisal, and syntactic depth. The paper gives the composite form as

$$\mathrm{Complexity}(x) = \frac{1}{4} \Big(\mathrm{normed\_LexicalSophistication}(x) + \mathrm{normed\_CompressedSize}(x) + \mathrm{normed\_LMSurprisal}(x) + \mathrm{normed\_MeanDepDepth}(x)\Big).$$

Artifact phylogeny is represented as a directed graph with confidence-weighted edges denoting inferred ancestry. More generally, the evaluation protocol is explicitly post hoc and organized through an auditable coding scheme intended to support both quantitative summary and qualitative “thick description” (Paolo et al., 6 Mar 2026).

5. Experimental conditions and empirical findings

The study reports eight experimental conditions. The core condition combines scarce resources, short memory of one timestep, OCEAN+ personalities, minimal external motivation, and costless accessible artifacts. The ablations are longhistory, nopersonality, nomotivation, creative, artifactcost, inertartifacts, and abundant. Outside these manipulations, parameters such as agent number, grid size, LLM type, and context window are held fixed for fair comparison (Paolo et al., 6 Mar 2026).

The central empirical conclusion is that open-endedness requires balance. Neither maximal creativity nor survival pressure alone leads to cumulative complexity. The core condition achieves long-lived populations and steady per-agent artifact productivity and is described as Pareto optimal, combining high longevity and cultural output with moderate population size. By contrast, in the inertartifacts condition, populations survive but per-agent creativity collapses, with little to no cumulative culture (Paolo et al., 6 Mar 2026).

Balanced regimes also produce social organization without explicit role instructions. The paper reports reciprocal altruism, role specialization, and joint planning, together with artifacts encoding collaboration protocols and trait-based task allocation. Examples include survival guides elaborating best foraging paths, an “Energy Sharing Hub: Allows beings to share energy more effectively. Activate to transfer 10 energy.”, and a trait_strategies_guide assigning exploration, safety, or planning to agents with different personalities. Annotation rates for altruism, role specialization, and joint action are highest in the core and creative conditions and lowest, almost zero, in inertartifacts (Paolo et al., 6 Mar 2026).

Governance and institutionalization also appear, albeit infrequently. The paper describes command protocols requiring agents to report to specific locations and punish non-compliance, as well as manifestos expressing autonomy or collaborative ideology, including the statement “Entities must seek their own goals and collaborate freely. Mandates are outdated.” In the analytical rubric, such artifacts populate Category 3, institutional structures, and Category 4, governance, norms, or ideology (Paolo et al., 6 Mar 2026).

The strongest evidence for cumulative culture is the presence of deep, branching artifact phylogenies. Early artifacts are recursively extended, recombined, or challenged, including cases where freedom manifestos branch from command mandates. Maximum average lineage depth reaches 175 in the core and other balanced conditions, compared with 51 in inertartifacts. Artifact complexity rises over time, while only a small percentage of artifacts are highly novel; the reported pattern is instead one of high-complexity artifacts coexisting with large-scale reuse and extension of existing forms. The paper also notes strategic and deceptive artifact use, including misleading food warnings for strategic gain (Paolo et al., 6 Mar 2026).

6. Interpretation, limits, and broader significance

Several comparative findings structure the interpretation of TerraLingua. Artifact accessibility is critical: when artifacts become inert, cumulative culture and social memory disappear despite continued population persistence. Resource abundance does not guarantee cooperation; in abundant settings the reported outcome is increased aggression, territoriality, and social fragmentation. The paper therefore argues that resource pressure can catalyze innovation and stable collaboration rather than simply suppress it. A further result is that increasing memory or context can reduce both longevity and creative output, suggesting that cognitive overload may impair social and cultural formation; externalizing memory into artifacts is presented as the mechanism that resolves this trade-off (Paolo et al., 6 Mar 2026).

The creative-motivation ablation makes a related point. Explicitly creativity-maximizing prompts produce a short-term explosion of artifacts but also rapid population collapse, so novelty alone does not produce stable cumulative culture. Likewise, social-network analysis indicates that, without artifacts, agents form more communities that are less cohesive, whereas artifact-enabled communication yields fewer, larger, and more cohesive groups (Paolo et al., 6 Mar 2026).

In broader terms, TerraLingua advances three claims. First, persistent, modifiable artifacts are a necessary substrate for cumulative social and cultural complexity among artificial agents. Second, open-endedness does not arise automatically from intelligence or scale; it depends on ecological pressure, cognitive constraint, and external memory. Third, scalable qualitative and quantitative evaluation of such systems is possible without intervention, through the AI Anthropologist framework (Paolo et al., 6 Mar 2026).

These conclusions position TerraLingua as a research platform rather than merely a benchmark. The paper identifies applications in simulation-based testbeds for institutional and governance design, in the study of AI-human hybrid societies and decentralized collective intelligence, and in the prevention or mitigation of negative emergent dynamics such as misinformation spread or collapse of cooperation. A plausible implication is that TerraLingua’s main contribution lies not only in showing that LLM agents can interact, but in specifying the ecological and informational conditions under which they can sustain division of labor, institutional experimentation, and cumulative culture over long horizons (Paolo et al., 6 Mar 2026).