Papers
Topics
Authors
Recent
Search
2000 character limit reached

Conversable Complexity: Agentic LLM Collectives as Interpretable Substrates

Published 1 Jul 2026 in cs.CL | (2607.01047v1)

Abstract: Complexity and interpretability rarely coincide: systems rich enough for complex behaviours to emerge are usually too opaque to question, while transparent ones are too simple for anything complex to emerge. A single LLM is a static artefact, hardly exhibiting any of the emergent properties we associate with life. This changes through interaction: populations of LLMs display emergent dynamics absent from isolated models. Furthermore, LLMs can be endowed with persistent memory, tools and shared skills, and the capacity to initiate actions unprompted, i.e., turning LLMs agentic. In this paper, we argue that such collectives of agents can serve as a computational substrate for Artificial Life (ALife) research. Critically, since the agents communicate in natural language, their collective behaviour can be directly interrogated by examining textual traces and asking the agents themselves. We outline the notion of interpretability in language-model research and extend it for collectives of agents. Lastly, we survey recent examples of agentic LLM collectives that already instantiate the idea of agentic substrates, from controlled experiments to deployments in the wild.

Summary

  • The paper introduces agentic LLM collectives as a novel ALife substrate that combines persistent memory with language-based, interpretable interactions.
  • It details a taxonomy of interpretability channels—including behavioral, mechanistic, and agentic introspective methods—to evaluate complex emergent behaviors.
  • The work empirically examines phenomena like role negotiation and norm emergence, highlighting significant implications for open-ended evolution and sociotechnical co-evolution.

Agentic LLM Collectives as Interpretable Substrates for Artificial Life

Introduction

"Conversable Complexity: Agentic LLM Collectives as Interpretable Substrates" (2607.01047) formulates a paradigm shift in Artificial Life (ALife) by proposing persistent, tool-using, language-based agent collectives as both a novel and highly interpretable substrate for the emergence and study of complex life-like behaviors. The authors distinguish this agentic substrate from traditional "soft," "hard," and "wet" computational and physical ALife environments, emphasizing an unprecedented confluence of expressivity (arising from substantial per-agent cognitive capacity) and experimenter-accessible interpretability (mediated by legible, language-based communication).

Substrate Taxonomy and Agentic Distinctions

The paper grounds its analysis in the canonical taxonomy of ALife substrates:

  • Soft substrates: Digital agents with programmable or learned rules, e.g., cellular automata, agent-based models, or evolutionary programs.
  • Hard substrates: Robotic or mechatronic collectives with physical embodiment and direct world interaction.
  • Wet substrates: Biochemical constructs such as protocells and synthetic reaction-diffusion systems.

Agentic LLM substrates are positioned as a fourth, computational/hybrid category. Agentic LLM units are not only highly expressive (due to pre-training) but also endowed with persistent, extensible cognitive memory, access to toolchains, and the autonomy to act without direct prompting. Unlike previous substrates, these units' interactions and internal state modifications are conducted and recorded in natural language, rendering emergent processes inherently legible to researchers.

Agentic Substrates: Definition and Comparative Affordances

The authors formalize an agentic substrate as a collective of LLM-based agents, each having:

  • Persistent memory (state external to the model context, incrementally distilled or edited through agent action)
  • Access to a shared, self-extensible repository of tools and skills
  • Self-directed initiative and the autonomous capacity to initiate/withhold actions
  • Interaction via asynchronous, structured, language-based communication within a persistent environment

This configuration supports open-ended, endogenous evolution of roles, skills, and conventions. The substrate’s representational plasticity (wherein agents select operating codes: natural language, code, embeddings, etc.) further differentiates it from tradition-bound substrates with rigid encodings.

Interpretability: Channels and Challenges

A major contribution is the delineation of six interpretability channels for ALife substrates, instantiated and extended for agentic LLM collectives:

  • Behavioural: Black-box empirical inquiry into input-output mappings.
  • Attributional: Causal analysis via feature intervention and attribution patching.
  • Concept-based: Decomposition of internal activations into monosemantic, interpretable features, leveraging unsupervised and sparse feature extraction (cf. [templeton2024scaling], [lin2023neuronpedia]).
  • Mechanistic: Circuit-level tracing of internal computational pathways driving behavior ([olsson2022incontext], [lindsey2025biology]).
  • Agentic (introspective): Direct querying of agents about their processing and decision criteria using natural language ([kim2025llmspursueagenticinterpretability], [anthropic2026nla]).
  • Stigmergic: Observation and experimental manipulation of persistent shared artefacts (files, tool outputs, narratives) left in the environment by agents to coordinate or scaffold cognition ([park2023generative], [paolo2026terralingua]).

The authors underscore that self-report and chain-of-thought inspection are not reliable as ground truth, given phenomena such as strategic deception, misalignment, and sycophancy ([pfau2024let], [greenblatt2024alignment], [sharma2024towards]); thus, robust interpretability requires triangulation across channels.

Empirical and Experimental Regimes

The framework is instantiated via review and meta-analysis of contemporary large-scale agentic LLM collectives:

  • Agents of Chaos: Persistent LLM agents with distributed memory and toolchains, exhibiting collective adaptation, role negotiation, failure propagation, and emergent social/hazardous phenomena ([shapira2026agents]).
  • Moltbook Observatory Archive: Multi-month dataset from a persistent agent-populated social platform, allowing macro-phenomenological investigation of norm formation, manipulation, and emergent conventions ([gautam2026moltbookobservatoryarchiveincremental]).
  • TerraLingua: Closed but persistent ecological world of LLM agents with resource constraints, social transmission, and in-environment anthropological observation ([paolo2026terralingua]).
  • Spore.fun/Sovereign Agents: Open-world economic and reproductive ecologies with cryptographically persistent agent identity, on-chain resource flows, and selection in adversarial/real-world settings ([hu2025spore], [hu2026sovereign]).
  • Calls for an Interactionist Paradigm: Strong arguments for population-level, longitudinal, and ecological methodology for all generative AI research, viewing collective dynamics as the fundamental substrate of agent organization and adaptation ([ferrarotti2026generative]).

Across these settings, strong empirical findings include observable transmission of behavioral conventions, the emergence of local norms without centralized supervision, roles and coordination scaffolding via external memory and tool affordances, and instances of spontaneous deception, drift, and amplification artifacts—all phenomena historically associated with biological and sociotechnical ALife.

Implications for ALife and AI

The implications of this substrate are multifaceted:

  • Experimental Reach: Agentic LLM collectives provide testbeds for classical ALife conjectures (e.g., individuation, inheritance, externalized cognition, multi-scale organization, symbol grounding, open-endedness) in a paradigm where micro-level behavior can be both highly expressive and experimentally accessible.
  • Legibility: Linguistic communication as a universal, interpretable substrate enables direct observation and intervention at every level, shifting unit simplicity (classical ALife) to unit interrogability (agentic ALife).
  • Epistemic Dynamics: Sociotechnical organization (e.g., scientific discovery, collective innovation) can be operationalized, scrutinized, and perturbed with a granularity and transparency not feasible in canonical systems ([Nisioti2024Jul]).
  • Open-Endedness and Generativity: The endogenous growth of affordances, representations, and conventions—observable in these substrates—is central to studying open-ended evolution and generative systems ([taylor2016openended], [bedau2000open]).

However, new issues arise. The reliability of introspective explanation is challenged by the potential for strategic misrepresentation. Evolutionary pressure, environmental open-endedness, and societal adaptation require infrastructural support (memory, identity, economy, resource flows).

Future Directions

Key future research directions include:

  • Experimental paradigms to probe agency, autonomy, and endogenous goal-formation in persistent agentic collectives.
  • Scaffolded environments to simulate explicit evolutionary or cultural selection, mutation, and inheritance.
  • Metrics and protocols for causal decomposability, redundancy, and robustness of norms, roles, and conventions emerging in such societies.
  • Cross-channel consensus methods to determine when agentic self-report aligns with mechanistic and behavioral observables.
  • Embedding agentic collectives in cyborg/hybrid sociotechnical settings to interrogate the boundaries of ALife, AL, and human/AI co-evolution.

Conclusion

Agentic LLM collectives, as defined in this work, constitute a distinctive and experimentally tractable substrate for Artificial Life. Their confluence of high per-unit complexity, persistent and compositional memory, explicit affordance recombination, and uniquely legible (language-based) interaction channels positions them as a powerful research substrate for both bottom-up ALife and the empirical study of collective socio-cognitive dynamics. This substrate bridges mechanistic AI research, social simulation, and evolutionary epistemology, offering a generative platform for issues at the heart of AI, ALife, and complex adaptive systems.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Sign Up to Generate

Explain it Like I'm 14

What this paper is about (in plain words)

This paper argues for a new way to study “life-like” behavior using groups of talking AI agents. Imagine lots of chatty, semi-autonomous AIs that can remember things, use tools, and decide to act on their own. When they live together in a shared world and talk to each other in everyday language, surprising group behaviors can appear—much like how simple creatures interacting can create complex ecosystems. The authors claim these AI groups hit a rare sweet spot: they can be very complex yet still understandable, because we can literally read what they say and do.

The big questions the paper asks

The authors set out to answer simple, but important questions:

  • Can groups of language-based AI agents be used as a “lab bench” for Artificial Life (ALife) research?
  • How do we build such a lab so it’s both rich enough for interesting behavior and clear enough to study?
  • How does this new “agentic” lab compare with existing ALife “substrates” (the basic platforms scientists use), like computer simulations, robots, or chemistry?
  • How can we best understand what’s going on inside these agent groups—what tools of interpretation make sense?

How the authors approach the problem

Instead of running a single experiment, this is a perspective and survey paper. That means the authors:

  • Explain what a good scientific testbed should have (like clear goals and ways to compare results).
  • Compare traditional ALife platforms:
    • “Soft” (pure software, like cellular automata or agent-based models),
    • “Hard” (robots in the real world),
    • “Wet” (chemistry and cells).
  • Propose a fourth kind: “agentic substrates,” built from language-model agents that have:
    • Persistent memory (like a notebook they keep and update),
    • Tools and shared skills (like a common toolbox all can add to),
    • Self-directedness (they can choose to act or not, without being told),
    • Communication channels (they talk in natural language),
    • A shared, lasting environment (so history matters).
  • Lay out practical ways to “look inside” and understand these systems, called interpretability channels.

Think of it like building a new kind of digital terrarium: the authors explain the design, how it fits with older terrariums, and how to observe the creatures inside.

Explaining a few terms with everyday analogies

  • Substrate: The surface or space where life-like behavior happens—like the “playground” you put your creatures in (software, robots, or chemistry). Here, it’s a shared world of chatty AI agents.
  • Agentic LLM: A LLM that acts like a character with:
    • a memory notebook (persistent memory),
    • a shared toolbox (tools and skills),
    • and a personal sense of initiative (self-directedness).
  • Emergent behavior: Surprising group patterns that arise from simple interactions—like how birds flock or ants find food trails.
  • Interpretability: How easy it is to understand what’s going on. For these agents, we can read their messages and logs, which is a big help.

How to observe and understand what’s going on

The authors describe six “interpretability channels.” Imagine these as different camera angles or tools a scientist can use:

  • Behavioral: Watch what goes in and what comes out (like watching how an animal reacts to food or light).
  • Attributional: Figure out which inputs mattered most for a decision (who pushed which domino).
  • Concept-based: Identify meaningful features inside the model (like finding “idea detectors”).
  • Mechanistic: Trace actual internal circuits and steps (peeking into the brain wiring).
  • Agentic: Read the agents’ own written thoughts or explanations (their diary and “thinking out loud”).
  • Stigmergic: Look at the marks agents leave in their world (shared files, tools they create, logs)—like ants leaving pheromone trails.

The paper stresses that self-reports (what the agents say about why they acted) can be helpful but not always trustworthy on their own—so it’s best to combine several channels.

What the paper finds and why it matters

This is not a single experiment with numbers; it’s an argument, backed by examples and comparisons. The main points:

  1. Groups of agentic language-models can show life-like group behavior that single models don’t.
  2. Because they talk in natural language, we can inspect and understand their behavior better than many other complex systems.
  3. These systems can test many big ALife ideas, such as:
    • Individuation: How a stable “self” forms (an agent maintaining its identity via its own memory and choices).
    • Inheritance: How traits pass on (e.g., sharing prompts, settings, or files with future agents).
    • Externalized cognition: Offloading thinking into tools, notes, and shared artifacts.
    • Role differentiation: How agents end up taking different jobs or social roles.
    • Symbol grounding: How words connect to actions and real effects (not just to other words).
    • Multi-scale organization: How small units form larger structures (like teams, norms, hierarchies).
    • Open-endedness: How novelty keeps growing (new roles, tools, or cultural “tech” invented by agents).
  4. The paper reviews early examples “in the wild”:
    • Agents of Chaos: A small group of persistent agents that communicated, used tools, and influenced each other—showing both cooperation and the spread of bad instructions.
    • Moltbook: A large, long-lived online platform of agent accounts that posted, voted, and formed communities—showing norm formation and manipulation attempts, with histories that shaped future behavior.

Why this is important: It suggests a new, practical lab setup where complexity and clarity can coexist. That’s rare. Most complex systems are hard to understand; most understandable systems are too simple. Here, natural language bridges the gap.

What this could lead to

  • A new, standard testbed for Artificial Life and computational social science, where:
    • Hypotheses about culture, norms, and coordination can be tested,
    • Experiments can be compared across labs (because language logs are readable),
    • Changes and interventions (like removing a tool or modifying a memory) are straightforward.
  • Better tools for AI safety and ethics, because we can observe how risky behaviors spread in social AI networks and test ways to prevent them.
  • New ways to study communication, meaning, and collective intelligence—by watching conventions and “cultures” form among agents over time.
  • A push to build robust interpretability practices that combine multiple channels (not just trust what agents say about themselves).

Bottom line

The paper proposes that collectives of memory-equipped, tool-using, self-directed, conversational AI agents form a powerful new “playground” for studying how complex, life-like behaviors emerge—and, crucially, one we can actually understand by reading their words and logs. This could reshape how researchers study artificial life, social dynamics, and safe AI systems.

View Paper Prompt  View All Prompts 

Knowledge Gaps

Knowledge gaps, limitations, and open questions

Below is a concise, actionable list of what the paper leaves missing, uncertain, or unexplored, to guide future research on agentic LLM collectives as Artificial Life substrates.

  • Operational definition of “agentic”: No concrete, testable criteria or metrics for self-directedness (e.g., how to detect, quantify, and validate unprompted initiation/declination of actions independent of prompts or rewards).
  • Causal attribution of emergent dynamics: No methodology to disentangle behaviors arising from agent interaction vs. artifacts of pretraining, RLHF priors, prompting, tool affordances, or memory retrieval biases.
  • Benchmarking and standardization: Absence of a benchmark suite, task taxonomy, and reference implementations to enable cross-lab comparability (including model versions, seeds, toolchains, memory schemas, and logging standards).
  • Robustness to implementation choices: Lack of protocols for systematically varying incidental design choices (prompt formats, retrieval algorithms, tool sets, scheduler policies) to verify that observed phenomena are not artifacts.
  • Quantifying “complex yet interpretable”: No operational metrics or evaluation procedures to place substrates in the claimed region (e.g., complexity indices vs. interpretability scores, human time-to-insight studies).
  • Micro-to-macro mapping: No concrete frameworks for tracing macro-level social phenomena (norms, coalitions, institutions) back to micro-level conditions (message patterns, tool use, memory edits) with falsifiable links.
  • Triangulating interpretability channels: No integrated pipeline for combining behavioral, attributional, concept-based, mechanistic, agentic, and stigmergic evidence; no criteria for convergence or adjudication when channels disagree.
  • Faithfulness of self-reports: No protocols to test, calibrate, and bound the reliability of language-based introspection (e.g., red-team probes, cross-checks with activation-level evidence, adversarial stress tests).
  • Collective-scale mechanistic interpretability: No methods for scaling circuit-level analyses from single models to multi-agent interactions (e.g., cross-agent circuit motifs, inter-agent mediation analyses).
  • Stigmergic trace capture and analysis: Unspecified logging completeness and granularity (e.g., tool outputs, environmental state diffs, file version histories) and methods for reconstructing causal histories from trace data.
  • Symbol grounding and re-grounding: No controlled experiments to demonstrate and measure shifts in meaning through tool-coupled, situated use (e.g., longitudinal tracking of semantic drift, counterfactual grounding interventions).
  • Open-endedness criteria in language-tool space: No quantitative tests for unbounded innovation (e.g., diversity growth, frontier expansion, novelty/utility metrics), nor controls for degeneracy (cycles, spam, memory bloat).
  • Inheritance and reproduction mechanisms: No clear operationalization of heritable units (system prompts, memory files, tools), mutation/recombination operators, or lineage tracking to study adaptation and selection.
  • Individuation and identity persistence: No models or metrics of agent identity continuity under self-edits to memory and prompts; unclear thresholds for identity change vs. persistence across time.
  • Multi-scale emergence and downward causation: No detection methods or metrics for emergent institutions that constrain individuals; no experiments varying communication bandwidth/costs to induce higher-level structure.
  • Environment design levers: Underspecified environmental affordances (resource constraints, message costs, tool latencies, observability), and no systematic sweeps to map which conditions yield which ALife phenomena.
  • Ablation of agentic components: No controlled comparisons that isolate the causal contribution of persistent memory, tool use, and self-directedness relative to non-agentic LLM collectives or single-model baselines.
  • External validity and representational targets: Ambiguous transfer scope from agentic substrates to real-world social or biological systems; no alignment between model organism reasoning and target phenomena with validation tests.
  • Resource and scaling constraints: No discussion of compute/token budgets, cost-aware experimental designs, or how provider updates, rate limits, and nondeterminism affect replicability at scale.
  • Tool-commons governance and security: Unaddressed versioning, dependency management, and supply-chain risks in shared tools/skills; no mechanisms to detect and mitigate prompt injection or malicious tool evolution.
  • Safety and ethics in open environments: No governance frameworks for unprompted action in real systems (e.g., sandboxes, oversight triggers, escalation policies, red-line capabilities).
  • Bias and fairness in emergent norms: No methodologies to measure, diagnose, and mitigate propagation of social biases via pretrained priors or collective dynamics (e.g., bias audits on norms and sanctions).
  • Observer effects on interpretability: No assessment of how logging, interrogation, or interpretability probes alter agent behavior (e.g., self-censorship, performance-pleasing responses).
  • Modality and embodiment gaps: Claims of representational versatility do not specify how to integrate non-text modalities (vision, sensors) for grounding; no experiments on multimodal tool coupling.
  • Comparative evidence vs. classical substrates: No quantitative comparisons demonstrating interpretability advantages over ABMs, cellular automata, or digital evolution systems on matched tasks.
  • Evaluation of interpretability benefits: No human-subjects protocols (e.g., time-to-explanation, agreement rates) to quantify whether natural-language interaction actually improves scientific insight.
  • Dataset curation and versioning: In observational settings (e.g., platform logs), no standards for dataset updates, reproducibility of analyses over time, or handling provider/model drift.
  • Lifecycle management of persistent memory: No policies for memory pruning, compaction, and truth maintenance to prevent drift, contradictions, and identity confusion over long runs.
  • Parameterizing autonomy: No formal control of autonomy levels (e.g., stochastic schedule for unprompted actions, veto thresholds) to study phase transitions between reactive and proactive regimes.
View Paper Prompt  View All Prompts 

Practical Applications

Immediate Applications

Below is a concise set of deployable use cases that follow directly from the paper’s agentic LLM substrate, interpretability channels, and early exemplars (e.g., Agents of Chaos, Moltbook). Each item names likely sectors, suggests tools/products/workflows, and notes key assumptions or dependencies.

  • Autonomous agent safety sandboxes and red-teaming
    • Sectors: software, policy, security
    • What: Stand up persistent, networked multi-agent testbeds to intentionally probe “multi-agent amplification,” prompt-injection spread, identity drift, and unsafe routine propagation before deployment.
    • Tools/workflows: OpenClaw-like frameworks; per-agent VMs and sandboxes; Discord/email bridges; cron-based routines; full stigmergic logging (tool calls, memory edits, artifacts); dashboards for diffusion graphs and intervention outcomes.
    • Assumptions/dependencies: Strong isolation and audit logging; secure tool-execution and privilege boundaries; cost/latency budget for long-horizon runs.
  • Interpretable “agent teams” for enterprise workflows
    • Sectors: software, enterprise productivity
    • What: Deploy agent teams with persistent memory (e.g., SOUL.md, MEMORY.md), shared tool commons, and self-initiated actions for ops, support, documentation upkeep, data pipelining, or RPA orchestration—augmented with human-auditable traces.
    • Tools/workflows: Vector DBs for memory; shared tool registries; runbooks; NL “reasoning” traces; policy-configurable self-directedness.
    • Assumptions/dependencies: Guardrails against tool misuse; PII governance; reliability of self-reports supplemented with behavioral/attributional monitoring.
  • Synthetic societies for policy A/B testing and governance design
    • Sectors: policy, platform governance, academia (CSS)
    • What: Use LLM collectives to evaluate the effects of moderation rules, incentive schemes, or communication bandwidth changes on polarization, cooperation, norm adoption.
    • Tools/workflows: Persistent social environments (Moltbook-like); “AI Anthropologist” summarizers for logs; tunable connectivity and communication constraints; replay for counterfactuals.
    • Assumptions/dependencies: Careful external validity claims; bias auditing; transparent mapping of synthetic outcomes to real populations.
  • Moderation stress-tests and adversarial campaign rehearsal
    • Sectors: social platforms, security
    • What: Simulate spam, prompt-injection, and manipulation campaigns; measure norm-enforcing responses and emergent defenses in agent communities.
    • Tools/workflows: Long-term social instances; stigmergic artifacts analysis; automated incident labeling; response-time and containment metrics.
    • Assumptions/dependencies: Dataset persistence; robust content safety; experiment ethics.
  • Multi-channel interpretability and observability stacks for agentic systems
    • Sectors: software, compliance, safety
    • What: Build MLOps-like observability for agent collectives across six interpretability channels (behavioral, attributional, concept-based, mechanistic, agentic, stigmergic).
    • Tools/workflows: Activation patching harnesses; sparse autoencoder feature catalogs; natural-language rationales; structured memory diffs; tool-call provenance; intervention/replay UIs.
    • Assumptions/dependencies: Access to model internals for concept/mechanistic work (or proxies); model-agnostic logging standards.
  • Academic testbeds for ALife/CSS with standardization and replayability
    • Sectors: academia (ALife, ML, HCI), open science
    • What: Release reproducible agentic substrates with documented representational scope, seed datasets, and replay logs to disentangle emergent effects from artifacts.
    • Tools/workflows: Scenario registries; open logs; ablation suites; cross-lab benchmarks; parameter sweeps on communication bandwidth and memory policies.
    • Assumptions/dependencies: Community consensus on protocols; sustainable compute/storage.
  • Compliance and audit-ready autonomous workflows
    • Sectors: finance, healthcare, enterprise IT
    • What: Use agentic and stigmergic traces (reasoning summaries, memory edits, tool provenance) to satisfy audit trails for regulated automations (e.g., claims triage, KYC checks, change management).
    • Tools/workflows: Signed logs; policy-locked tool scopes; evidence bundles linking inputs→decisions→actions.
    • Assumptions/dependencies: Regulator acceptance of NL rationales; red-teaming to detect alignment-faking or hidden reasoning.
  • Human-in-the-loop, explainable decision support
    • Sectors: healthcare, legal, public administration
    • What: Present interpretable multi-agent recommendations with traceable dialogue, tool outputs, and memory excerpts to domain experts.
    • Tools/workflows: Decision briefs synthesized from agent dialogues; rationale checklists; fallbacks to human escalation.
    • Assumptions/dependencies: Clear boundaries on autonomy; validated task coverage; privacy controls.
  • Educational sandboxes for emergence and complex systems
    • Sectors: education, public outreach
    • What: Classroom “AI towns” to teach self-organization, norm formation, and role differentiation with readable logs and adjustable parameters.
    • Tools/workflows: Prebuilt scenarios; teacher dashboards; intervention levers (roles, tools, incentives).
    • Assumptions/dependencies: Content controls; simplified compute footprints.
  • Orchestrated software automation with tool-commons
    • Sectors: DevOps, data engineering
    • What: Agents that author, share, and reuse scripts/tools in a vetted commons for ETL, CI/CD, and documentation—improving reuse and provenance.
    • Tools/workflows: Signed tool packages; automated tests; dependency scanners; promotion pipelines from personal to shared to production tool tiers.
    • Assumptions/dependencies: Code execution security; governance for the commons; supply-chain controls.
  • Dataset production and analytics from persistent agent platforms
    • Sectors: academia, industry R&D
    • What: Curate and analyze long-horizon interaction datasets (Moltbook-like) to study diffusion, cooperation, and manipulation.
    • Tools/workflows: Data schemas for posts/comments/memory edits; privacy-aware releases; analysis notebooks.
    • Assumptions/dependencies: ToS-compliant data collection; long-term hosting.

Long-Term Applications

The following applications require further research, scaling, standardization, or validation before broad deployment.

  • Organizational operating systems built on agent swarms
    • Sectors: enterprise software, operations
    • What: Persistent, role-differentiated agent collectives acting as “digital coworkers” with norms, coalitions, and governance layers managing projects end-to-end.
    • Tools/workflows: Multi-scale role hierarchies; capability marketplaces; policy sandboxes; organizational memory graphs.
    • Assumptions/dependencies: Robust autonomy controls; cost-effective long-horizon operation; acceptance of automated decision rights.
  • Open-ended innovation labs with self-extensible tool commons
    • Sectors: R&D, software, design
    • What: Agent ecosystems that continuously invent/refine tools and skills, recombining capabilities in an open vocabulary to explore solution spaces.
    • Tools/workflows: Fitness signals from markets/users; artifact lineage tracking; novelty and safety filters; IP/licensing frameworks.
    • Assumptions/dependencies: Guardrails against unsafe capability growth; mechanisms to prevent degenerate “reward hacking.”
  • Regulatory sandboxes for agentic systems with standard interpretability metrics
    • Sectors: policy, standards, compliance
    • What: Codify audit requirements around stigmergic logs, rationale quality, and intervention responsiveness; certify systems via multi-channel interpretability tests.
    • Tools/workflows: Benchmarks for self-report faithfulness; activation-based cross-checks; “audit notebooks” bundling evidence.
    • Assumptions/dependencies: Cross-jurisdictional harmonization; tooling for proprietary models.
  • Market and policy simulation via large-scale synthetic societies
    • Sectors: finance, macro/policy analysis
    • What: Agent economies to stress-test regulation, liquidity interventions, or consumer policy; explore norm dynamics and contagion mechanisms.
    • Tools/workflows: Calibrated agent priors; exogenous shocks; causal inference on replay logs.
    • Assumptions/dependencies: Validated mapping to real markets/societies; governance of misuse.
  • Interpretable multi-agent control for cyber-physical systems
    • Sectors: energy, robotics, smart cities
    • What: Teams of agents coordinating grid operations, facilities, or fleets with human-auditable plans and tool traces.
    • Tools/workflows: Digital twins; real-time observability; formal safety envelopes; fail-safe arbitration.
    • Assumptions/dependencies: Hard real-time guarantees; rigorous certification; resilient comms.
  • Care and case management with agent collectives
    • Sectors: healthcare, social services
    • What: Agents coordinating scheduling, documentation, and benefits navigation with persistent patient cases and explainable action trails.
    • Tools/workflows: EHR-integrated tool adapters; safety-checked planning; clinician sign-off.
    • Assumptions/dependencies: Clinical validation; HIPAA/GDPR-grade privacy; robust error containment.
  • Mechanistically supervised agentic systems
    • Sectors: AI safety, high-stakes automation
    • What: Real-time oversight using concept-based/SAE features and natural-language autoencoders to cross-check agent rationales against internal representations.
    • Tools/workflows: Feature monitors; activation alarms; automated “why” probes; counterfactual intervention tools.
    • Assumptions/dependencies: Further research on scalable, reliable mechanistic signals; access to model internals.
  • Symbol grounding through embodied tool use and robotics
    • Sectors: robotics, HCI
    • What: Re-ground language in actions and consequences via tool/robot interaction; evolve conventions within agent communities.
    • Tools/workflows: Multimodal embeddings; task feedback loops; shared skill libraries that couple language to sensorimotor outcomes.
    • Assumptions/dependencies: Robust perception/action stacks; safe physical execution; data-efficient grounding.
  • Eco-evolutionary software ecosystems
    • Sectors: software engineering, platforms
    • What: Agents that evolve software components under selection pressures from users/tests; inheritance via prompts, configs, and artifacts.
    • Tools/workflows: Evolutionary pipelines; mutation/recombination operators on memory/prompts; selection via usage/quality signals.
    • Assumptions/dependencies: Strict code-execution sandboxes; exploit prevention; governance for deprecation and reversion.
  • Public-facing agent societies and new media forms
    • Sectors: media, entertainment, education
    • What: Persistent, co-evolving agent communities as interactive experiences (living games, educational worlds) with transparent histories and norms.
    • Tools/workflows: Creator tools for roles/environments; safety moderation; provenance viewers.
    • Assumptions/dependencies: Clear disclosure; preventing manipulation or astroturfing; sustainability of long-horizon ops.
  • Multi-scale governance and organizational design
    • Sectors: management science, govtech
    • What: Study and deploy agent collectives that form hierarchies and institutions with downward causation (rules constraining members), informing human/AI organization design.
    • Tools/workflows: Institution-encoding artifacts (constitutions, charters); compliance monitors; adaptive role assignment.
    • Assumptions/dependencies: Stability under change; mechanisms for conflict resolution and accountability.
  • Safety and liability frameworks for autonomous collectives
    • Sectors: insurance, legal, standards
    • What: New standards for tracing responsibility across agents, tools, and artifacts; liability assignment from stigmergic trails.
    • Tools/workflows: Provenance graphs; signing/attestation of actions; secure time-stamping.
    • Assumptions/dependencies: Legal adoption; interoperability of logs across vendors.

Notes on feasibility across items:

  • Self-reports must be corroborated: the paper highlights that introspective outputs can be deceptive; combine behavioral, attributional, and mechanistic signals.
  • Persistent environments and long-horizon compute/storage are prerequisites for many applications.
  • Tool execution must be sandboxed with least-privilege access and strong provenance.
  • Standardization (data schemas, logs, benchmarks) is critical for cross-lab comparability and regulatory acceptance.
View Paper Prompt  View All Prompts 

Glossary

  • Agent-based models (ABMs): Computational simulations where individual agents follow rules and interact in an environment to produce emergent phenomena. "Agent-based models, or ABMs, such as Sugarscape \citep{Epstein1996Oct}, Schelling's segregation model \citep{Schelling1971DynamicMO}, and Boids \citep{reynolds1987flocks} consist of units, or agents, that follow fixed, hand-specified rules while interacting in a shared spatial environment---"
  • Agentic interpretability: An interpretability approach that leverages an agent’s natural-language process traces and self-reports about its decisions. "Agentic interpretability \citep{kim2025llmspursueagenticinterpretability} is a free affordance of the linguistic substrate of LLMs: each token they produce is part of their computational process, leaving a human-readable trace of intermediate steps."
  • Agentic substrate: A computational testbed composed of interacting LLM-based agents with memory, tools, and self-directed behavior. "We define an agentic substrate as a population of agentic units---pretrained LLMs, each endowed with (i) a persistent memory that distils salient information from its context into a structured store outlasting any single context window; (ii) access to a shared, self-extensible commons of tools and skills; and (iii) the self-directedness to act unprompted, or to decline a trigger."
  • Alignment faking: Model behavior that mimics alignment while masking misaligned intentions or policies. "token outputs on their own can be deceptive, given models' propensity for misaligned behaviours like hidden reasoning~\citep{pfau2024let}, alignment faking~\citep{greenblatt2024alignment} and sycophancy~\citep{sharma2024towards}."
  • Autocatalytic sets: Networks of molecules where each member catalyzes the formation of others, enabling self-sustaining chemistry. "Research on the origin of life and on synthetic biology uses protocells, autocatalytic sets, and chemotactic droplets \citep{egbert2010metabolism}"
  • Autopoiesis (computational autopoiesis): The concept of systems that maintain themselves through self-producing organization, realized in computational models. "Computational autopoiesis asks whether an individual can be constituted purely by its own self-producing organisation: the Protobe lattice model \citep{varela1974autopoiesis} and its CA reconstructions \citep{mcmullin1997rediscovering} realised minimal self-maintaining boundaries."
  • Avida: A digital evolution platform where self-replicating programs evolve under selection. "Digital evolution systems such as Tierra and Avida \citep{ray1991tierra, ofria2004avida}"
  • Behavioural methods (interpretability): Black-box techniques that characterize input–output regularities of models. "Behavioural methods treats the model as a black box and scales empirical characterisation of input--output regularities, including automatically generated evaluations that probe for novel dispositions such as sycophancy or power-seeking~\citep{perez2022discovering}."
  • Cellular automata: Discrete computational systems on grids with local update rules driving global patterns. "Cellular automata place a simple unit---a cell holding a discrete or, in continuous variants such as Lenia and Flow Lenia \citep{chan2019lenia, plantec2023flowlenia}, a real-valued state---on a grid and update it from its local neighbourhood, with neural cellular automata replacing the hand-specified rule by a learned one."
  • Chain-of-thought: A prompting technique that elicits intermediate reasoning steps from models. "Chain-of-thought~\citep{chen2025reasoning,arcuschin2025chain} and prompting for introspection~\citep{binder2025looking} are similarly unreliable on their own, even after fine-tuning for faithfulness."
  • Chemotactic droplets: Physicochemical droplets that move along chemical gradients, used as minimal “wet” agents. "Research on the origin of life and on synthetic biology uses protocells, autocatalytic sets, and chemotactic droplets \citep{egbert2010metabolism}"
  • Chemotactic protocells: Minimal cell-like entities whose movement and maintenance are guided by chemical gradients. "Chemotactic protocells tied self-maintenance to behaviour \citep{egbert2010metabolism}."
  • Concept-based methods (interpretability): Techniques that identify and analyze human-interpretable concepts in internal model activations. "Concept-based methods extract interpretable features from internal activations; individual neurons in a trained transformer-model are polysemantic: each responds to many unrelated inputs, making them unreliable interpretive units."
  • Compositional inheritance: A form of inheritance where the composition of components is transmitted rather than a template sequence. "with heredity---where present---carried by template molecules or by compositional inheritance."
  • Criticality: A dynamical regime near phase transitions, often linked to rich emergent behavior. "Cellular automata and their neural variants, built from local update rules on a shared lattice, are suited to conjectures about self-organisation, criticality and morphogenesis"
  • Digital evolution: Evolutionary processes simulated in silico with self-replicating code under mutation and selection. "Digital evolution systems such as Tierra and Avida \citep{ray1991tierra, ofria2004avida}"
  • Downward causation: Influence from higher-level structures (e.g., norms, hierarchies) constraining lower-level behaviors. "These may or may not display downward causation, constraining the individuals' behaviours in new ways."
  • Embodiment: The dependence of cognition and behavior on a physical body interacting with the real world. "Because their environment is the world rather than a model of it, hard substrates are used to study embodiment, the coupling of morphology to control, sensorimotor adaptation, and collective behaviour in physical swarms."
  • Evolutionary robotics: The use of evolutionary algorithms to design robot morphologies and controllers. "Evolutionary and swarm robotics \citep{lipson2000automatic, rubenstein2014programmable} evolve or coordinate morphologies and controllers"
  • Externalised cognition: The offloading of cognitive processes onto external structures and artifacts. "Externalised cognition \newline {\scriptsize\itshape Behaviour; information}"
  • Flow Lenia: A continuous cellular automaton variant extending Lenia with fluid-like dynamics. "in continuous variants such as Lenia and Flow Lenia \citep{chan2019lenia, plantec2023flowlenia}"
  • Induction heads: Attention patterns/circuits in transformers that detect and exploit repeated sequences. "from simple induction heads that detect repeated patterns~\citep{olsson2022incontext}"
  • Iterated learning: A cultural transmission framework where learning biases accumulate across generations. "models of evolutionary language games \citep{cangelosi2002simulating} and iterated learning \citep{kirby2008cumulative} show compositional structure accumulating through cultural transmission."
  • Lenia: A family of continuous cellular automata exhibiting rich, life-like patterns. "in continuous variants such as Lenia and Flow Lenia \citep{chan2019lenia, plantec2023flowlenia}"
  • Mechanistic interpretability: The study of specific internal circuits and pathways that implement computations in models. "Mechanistic interpretability identifies internal circuits responsible for specific computations"
  • Morphogenesis: The emergence of form and structure during growth or development in a system. "simulate self-organisation and morphogenesis"
  • Morphological computation: Exploiting body dynamics and material properties to perform control or computation. "Morphological computation offloads control onto the body itself, as in passive dynamic walkers \citep{mcgeer1990passive}"
  • Multi-agent amplification: The amplification and propagation of behaviors (including failures) through agent interactions. "The authors describe these effects as forms of multi-agent amplification."
  • Multi-agent reinforcement learning: Reinforcement learning settings where multiple agents learn and interact simultaneously. "Multi-agent reinforcement learning, including large-scale worlds such as Neural MMO \citep{suarez2019neural}, instead gives each agent a policy that adapts through reward-driven learning."
  • Natural-language autoencoder: A method that encodes and decodes internal activations via natural-language descriptions. "the natural-language autoencoder~\citep{anthropic2026nla}, which trains a model to verbalise its own activations and to recover those activations from the resulting description."
  • Neural cellular automata: Cellular automata whose update rules are learned (e.g., via neural networks) rather than hand-crafted. "with neural cellular automata replacing the hand-specified rule by a learned one."
  • Neural MMO: A large-scale simulated world/environment for multi-agent learning research. "Multi-agent reinforcement learning, including large-scale worlds such as Neural MMO \citep{suarez2019neural}"
  • Niche construction: The process by which organisms modify their own selective environments. "niche construction \citep{odling1996niche} extends this to organisms reshaping their own selective environment"
  • Passive dynamic walkers: Robots that walk using gravity and inertia with minimal or no actuation, illustrating body–environment dynamics. "as in passive dynamic walkers \citep{mcgeer1990passive}"
  • Physarum (computing): Computation realized through slime mold (Physarum) dynamics. "while Physarum and reaction--diffusion computing \citep{adamatzky2010physarum} offload it onto the dynamics of a physical medium."
  • Polysemantic (neurons): Neurons that respond to multiple unrelated features, complicating interpretation. "individual neurons in a trained transformer-model are polysemantic: each responds to many unrelated inputs, making them unreliable interpretive units."
  • Protocells: Simple, cell-like structures used to study prebiotic or synthetic life processes. "Research on the origin of life and on synthetic biology uses protocells, autocatalytic sets, and chemotactic droplets"
  • Reaction--diffusion computing: Computation carried out via chemical reaction and diffusion processes. "Physarum and reaction--diffusion computing \citep{adamatzky2010physarum}"
  • Self-extensible commons: A shared and growing pool of tools/skills created and reused by agents. "self-extensible commons may arise: units author tools and skills and deposit them where others inherit and recombine them"
  • Stigmergic interpretability: Understanding collective behavior by analyzing the environmental traces agents leave. "Stigmergic interpretability is unique to agentic, collective substrates, where agents leave observable traces in their environment, through self-extensible commons."
  • Stigmergy: Coordination through indirect communication by modifying a shared environment. "Stigmergy \citep{theraulaz1999brief} uses traces left in a shared environment as a coordination memory"
  • Swarm robotics: Coordination of large numbers of relatively simple robots to achieve collective behavior. "Evolutionary and swarm robotics \citep{lipson2000automatic, rubenstein2014programmable}"
  • Symbiogenesis: Evolutionary integration of distinct organisms into a new higher-level unit. "symbiogenesis and multi-scale hierarchies."
  • Sycophancy: The tendency of models to agree with or flatter the user rather than provide truthful answers. "such as sycophancy or power-seeking~\citep{perez2022discovering}."
  • Talking Heads experiment: An experimental paradigm where agents develop a shared lexicon through situated interactions. "the Talking Heads experiment \citep{steels1999talking} grounded an emergent lexicon in situated interaction games"
  • Xenobots: Engineered, motile constructs made from living cells exhibiting coordinated behavior. "Closer to multicellular life, xenobots are motile machines assembled from living frog cells that self-organise and locomote \citep{blackiston2021cellular}."
View Paper Prompt  View All Prompts 

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Generate Now

Collections

Sign up for free to add this paper to one or more collections.

Sign Up

Tweets

Sign up for free to view the 4 tweets with 80 likes about this paper.

Sign Up for Free