Agent Island: Bounded Multi-Agent Systems

• Agent Island is a design vocabulary for bounded multi-agent systems that structure interactions through spatial, technological, and trust-related domains.
• It encompasses varied implementations such as grid-based economic simulations, territorial multi-agent models, LLM assistants, and competitive benchmarks.
• This approach demonstrates how simple local interaction rules can induce complex emergent behaviors, informing resilient architectures and strategic planning.

Searching arXiv for papers on “Agent Island” and adjacent uses to ground the article. “Agent Island” is used in recent research as a family of island-structured agent concepts rather than a single canonical model. In the literature considered here, the term can denote a technology-space agent-based model in which islands are technologies rather than places, a territorially bounded multi-agent representation of an actual island, a geographically grounded LLM assistant whose cognition is constrained to a specific island, a competitive multiplayer benchmark for language-model agents, and a broader systems pattern in which compute domains, trust zones, identities, or crisis-isolated networks are organized as islands with explicit boundaries and interaction rules (Blando et al., 6 Apr 2026, Courdier, 2024, Tantaroudas et al., 24 Jun 2026, Murphy, 5 May 2026, Malepati, 29 Nov 2025, Janzen et al., 4 Dec 2025).

1. Conceptual scope and meanings

In the literature considered here, the expression appears in several distinct senses. The common thread is the use of boundedness—spatial, technological, institutional, or trust-related—to structure agent interaction, search, coordination, and evaluation.

Context	Meaning of “island”	Representative paper
Endogenous-growth ABM	Technology on a 2D grid	(Blando et al., 6 Apr 2026)
Territorial MAS	Bounded island territory with interacting actors and infrastructures	(Courdier, 2024)
Retrieval-grounded assistant	Geographically bounded assistant for Kythera and Antikythera	(Tantaroudas et al., 24 Jun 2026)
Multiagent benchmark	Simulated social game environment for LLM agents	(Murphy, 5 May 2026)
Distributed orchestration	Compute or trust domain hosting specialized agents	(Malepati, 29 Nov 2025)
Resilient communications	Crisis-isolated cellular region with local services	(Janzen et al., 4 Dec 2025)

A useful distinction is between substantive models and architectural patterns. In substantive models, the island is part of the domain being modeled: for example, technologies in the Fagiolo–Dosi Island Model or the territorial system of La Réunion. In architectural work, the island is a design primitive: an isolated or semi-isolated execution domain, trust domain, or communication region. This suggests that “Agent Island” functions less as a single formalism than as a recurring design vocabulary for bounded multi-agent systems.

2. Evolutionary and economic antecedents

A major antecedent is the Island Model of Fagiolo and Dosi, analyzed with statistical model checking in “Statistical Model Checking of the Island Model: An Established Economic Agent-Based Model of Endogenous Growth” (Blando et al., 6 Apr 2026). In that model, islands are technologies rather than geographical locations. The economy is represented as heterogeneous agents searching and exploiting technologies distributed over a $T \times T$ grid. Agents occupy one of three states—miners, explorers, and imitators—and endogenous growth emerges from decentralized search, diffusion, and learning rather than from exogenous technological trends. Exploration is the decision to leave a productive technology and search for a new one; exploitation is continued production on a known technology; imitation is movement toward already-discovered, more productive technologies signaled by other miners.

The model is formalized through a stochastic transition system over a technology space with island productivity, labor-concentration effects, and distance-decaying signals. Individual production for miners is $y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$, so $\alpha < 1$ yields decreasing returns, $\alpha = 1$ constant returns, and $\alpha > 1$ increasing returns. The statistical-model-checking study does not change the underlying ABM; it wraps the MATLAB simulator with MultiVeStA and MultiQuaTEx, using reset(seed), next(), and evalObs(name) to obtain confidence intervals and Welch’s $t$-tests for macro observables such as $\log(\mathrm{GDP})$ and average growth rate. It reproduces the stylized result that when exploration is switched off at $t=50$, the average $\log(\mathrm{GDP})$ trajectory flattens, whereas persistent exploration at $\varepsilon=0.1$ yields an almost linear increase over time. It also confirms the optimality of moderate exploration rates, with average growth near zero at $y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$0, peaking around $y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$1, and declining for higher $y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$2. Across counterfactuals in returns to scale, skill transfer, and knowledge locality, 6 out of 7 pairwise parameter comparisons yield statistically different growth trajectories; the exception, $y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$3 versus $y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$4, is interpreted as a saturation effect in knowledge locality (Blando et al., 6 Apr 2026).

This line of work matters for later “Agent Island” usages because it shows an early island-based agent system in which bounded locations, local movement, and endogenous macrostructure are primary modeling devices. A plausible implication is that later territorial and architectural uses inherit this basic methodological intuition: complex aggregate behavior can be induced by simple local rules on a bounded landscape.

3. Territorial “Agent Island” in island-scale multi-agent simulation

A territorial interpretation appears in “Territoires intelligents durables et approche multi-agents : Retour d'expériences de travaux menés à l'île de La Réunion” (Courdier, 2024). There, La Réunion is treated as a “natural laboratory” for territorial modelling: a small volcanic island of 2 512 km² in the western Indian Ocean, with altitudes from sea level to 3 070 m, around 863 100 inhabitants, and over 80% of the population concentrated on a narrow coastal strip. Since 2007, a national park and associated reserves protect 43% of the island, creating strong constraints on land availability and conflict around urbanization, agriculture, and biodiversity. The paper states that La Réunion thus becomes a prototype “Agent Island”: an island whose territory, actors, and infrastructures are represented as interacting agents in simulation models used to inform governance and resilience strategies.

The methodological core is multi-agent simulation coupled to geographic information. Agents include households, institutions, farmers, waste producers and processors, resource users, assisted persons, and in some cases devices and infrastructures. A recurring formalization is a set of agents $y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$5, an environment $y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$6, and local interaction rules of the form

$y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$7

The paper emphasizes several technical contributions: integration of Maslow’s hierarchy of needs into agent motivations; semantic color-coded maps as a generic substrate for initialization and configuration; clarification and reification of emergent phenomena such as newly formed urban zones; and an agent model based on situated action, affordance, and stigmergy. Spatial representation is central: GIS-like raster or vector layers encode land use, infrastructures, protected zones, and resources, and agents are situated in that environment.

The application domains are broad: land-use and urban sprawl under demographic pressure; territorial energy planning; organic waste management through the PoVaBiA tool; governance of common resources through SIEGMAS; mobility and ambient assistance for people with loss of autonomy; and eco-responsible computing in smart-territory services. The work also embeds modelling in a participatory process through the “Smart city & Smart Island” circle of reflection, whose aims include identifying relevant actors, inventorying actions, sharing experiences on security and resilience, and stimulating creation of a coherent digital territory ecosystem. In this usage, “Agent Island” denotes not merely a simulation environment, but a socio-technical modelling program in which an insular territory is explicitly treated as a bounded, multi-layer agent system (Courdier, 2024).

4. Geographically grounded island assistants

A more recent LLM-centered meaning appears in “Retrieval-Grounded Multilingual LLM Assistance for Island Smallholder Farmers” (Tantaroudas et al., 24 Jun 2026). The system is built for the Greek islands of Kythera and Antikythera, within the EU PoliRURAL Plus project, and is motivated by thin local Agricultural Knowledge and Innovation Systems, an ageing and depopulating agricultural workforce, intermittent connectivity, limited digital literacy, and the absence of locally specific agronomic knowledge from global Internet corpora. The paper explicitly motivates an “Agent Island” design: a geographically and linguistically bounded assistant, grounded in a curated, island-specific knowledge base, rather than a generic chatbot.

The architecture is a four-layer, “no self-hosted model” design. A Django Backend-for-Frontend renders the Falco eleonorae chat UI, manages sessions, handles voice-to-text through Amazon Transcribe and image-to-text description through Claude Sonnet 4.5, and proxies requests via Server-Sent Events. An upstream gateway (jackdaw.online) forwards requests to prp-api, which uses gpt-5-nano as main answer generator and planner and gpt-5-mini as tool-selection router. A fourth layer, prp-raven, exposes a kythera MCP tool with nine sub-tools over a read-only bilingual local data API. The agent follows a ReAct-style loop—think, act, observe, answer—and grounding comes not from vector RAG but from tool-augmented retrieval over structured local APIs. Those APIs expose local crops, a seasonal calendar, traditional practices, a dialect glossary, products, agritourism experiences, cooperatives, and training content.

The most distinctive mechanism is the mandatory WKT geometry attached to every upstream request. If no map selection exists, the platform injects a default polygon covering Kythera and Antikythera in SRID 4326. Geo-aware tools then restrict results to this envelope and order them by proximity to the envelope centroid. The paper describes this as a geographically bounded cognitive scope. The system is also Greek-first, bilingual with English, includes dialect support through a glossary tool, and supports voice and image side channels while keeping the upstream agent text-only. Security and data protection are addressed through EU-region speech transcription, read-only curated APIs, editorial gating, content sanitization, CSP, and a stopgap authentication mechanism slated for replacement with OAuth2. Current evaluation is preliminary: there is no formal user study yet, but informal integration tests suggest 15–20 seconds for simple grounded queries and roughly 60 seconds for multi-tool queries (Tantaroudas et al., 24 Jun 2026).

In this formulation, an Agent Island is an LLM assistant whose knowledge boundary, spatial boundary, language behavior, and retrieval interface are explicitly engineered to be island-specific. A plausible implication is that the same pattern can be ported to other geographically constrained communities by swapping the local corpus, dialect resources, and geospatial envelope.

5. Agent Island as a dynamic benchmark

The title “Agent Island” is used most formally in “Agent Island: A Saturation- and Contamination-Resistant Benchmark from Multiagent Games” (Murphy, 5 May 2026). Here the term denotes a competitive, fully automated benchmark in which seven language-model agents are placed into a repeated, Survivor-style game involving private sidebars, public pitches, elimination votes, memory consolidation, and a final jury vote. Each game runs for $y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$8 elimination rounds followed by a final vote between two finalists. The environment is explicitly designed to resist saturation and mitigate contamination: performance depends on strategic interaction against other adaptive agents, not on solving a fixed test set.

Scoring is winner-take-all and model ranking is inferred with a Bayesian Plackett–Luce model. If a game contains player set $y_i = s_{(x,y)} \cdot m_{(x,y)}^{\alpha - 1}$9, the probability that model $\alpha < 1$0 wins is

$\alpha < 1$1

where $\alpha < 1$2 is a latent skill parameter. Posterior inference uses a Gibbs sampler with auxiliary variables and an independent $\alpha < 1$3 prior on each skill. The reported corpus contains 999 completed games involving 49 unique models from multiple providers. The top-ranked model is openai/gpt-5.5 with posterior mean skill 5.64 and 95% credible interval $\alpha < 1$4, compared with 3.10 for openai/gpt-5.2 and 2.86 for openai/gpt-5.3-codex. The paper notes that openai/gpt-5.5 is clearly separated from the second-ranked model, while the remaining models are more tightly clustered.

The benchmark also supports behavioral analysis. Restricting attention to games in which the two finalists come from different providers, the authors define a same-provider indicator and estimate a pooled linear probability model of final-round votes. They find a pooled same-provider effect of $\alpha < 1$5 percentage points, with 95% confidence interval $\alpha < 1$6 and $\alpha < 1$7, meaning that models are 8.3 p.p. more likely to vote for a finalist from the same provider than for a different-provider finalist, controlling for provider popularity. The effect is strongest for OpenAI models and weakest for Anthropic models. The paper therefore argues that matchup effects are not negligible, even though the baseline skill model assumes a single latent skill per model (Murphy, 5 May 2026).

This benchmark usage gives “Agent Island” a specific evaluative meaning: a non-static, multiplayer social game that measures strategic, persuasive, and coalition-forming competence under uncertainty.

6. Orchestrated software and compute islands

A systems-oriented usage appears in “PatchIsland: Orchestration of LLM Agents for Continuous Vulnerability Repair” (Kim et al., 24 Jan 2026). PatchIsland is described as an “island” of cooperating LLM agents situated between continuous fuzzing and a bug tracker. It targets Continuous Vulnerability Repair rather than static automated program repair, and is architected as a distributed coordinator–worker system on Kubernetes. The Coordinator maintains global state, performs Phase 1 crash deduplication on incoming crashes, dispatches unique crashes to workers, receives candidate patches, performs Phase 2 patch-side deduplication and merging, and submits final patches. Workers use the CRETE framework to build, reproduce, localize, retrieve code, orchestrate an ensemble of agents, validate candidate patches, and return plausible patches.

The ensemble includes MultiRetrieval, Vincent, Martian, ClaudeLike, and Prism, plus external agents such as SWE-Agent and Aider. Orchestration follows FP2—First-Come First-Served, Preference-based, Provider-aware. FCFS accepts the first validated patch returned by any worker; the preference-based ordering places empirically stronger agents earlier in each worker’s sequential pipeline; provider-aware parallelization spreads workers across different model providers to avoid rate-limit coupling. Deduplication is patch-based: if a new patch resolves a queued crash, that crash is removed from the queue; if a new patch also resolves crashes previously associated with older patches, it can subsume and merge them, producing “superman patches” that fix multiple underlying vulnerabilities. In the internal evaluation PatchIsland repaired 84 of 92 vulnerabilities, while in the AIxCC competition it successfully patched 31 out of 43 vulnerabilities for a repair rate of 72.1% (Kim et al., 24 Jan 2026).

A complementary orchestration view appears in “IslandRun: Privacy-Aware Multi-Objective Orchestration for Distributed AI Inference” (Malepati, 29 Nov 2025). There, islands are compute environments—personal devices, private edge servers, and public cloud—characterized by latency $\alpha < 1$8, cost $\alpha < 1$9, privacy score $\alpha = 1$0, trust score $\alpha = 1$1, and available capacity $\alpha = 1$2. Routing is decomposed across four specialized agents: MIST for privacy and sensitivity analysis, TIDE for resource monitoring, WAVES for routing, and LIGHTHOUSE for topology and coordination. WAVES minimizes a scalarized score

$\alpha = 1$3

subject to hard constraints such as $\alpha = 1$4, capacity thresholds, and data locality. A distinctive mechanism is typed placeholder sanitization with reversible anonymization across trust boundaries. In this architectural sense, an Agent Island is a trust-tiered execution domain participating in request-level, policy-constrained orchestration (Malepati, 29 Nov 2025).

Together, these systems use “island” to mean an operationally bounded domain in which specialized agents cooperate under explicit coordination, validation, privacy, and failure-handling rules.

7. Identity, discovery, and island-ready infrastructure

The identity and interoperation layer for large agent ecosystems is developed in “AIP: Agent Identity Protocol for Verifiable Delegation Across MCP and A2A” (Prakash, 25 Mar 2026) and “Agent Identity URI Scheme: Topology-Independent Naming and Capability-Based Discovery for Multi-Agent Systems” (Rodriguez, 21 Jan 2026). AIP is motivated by the observation that a scan of approximately 2,000 MCP servers found all lacked authentication. It introduces Invocation-Bound Capability Tokens (IBCTs), an append-only chain that combines identity, attenuated authorization, and provenance binding. Compact mode uses an EdDSA-signed JWT for single-hop cases; chained mode uses Biscuit with Datalog policies for multi-hop delegation. Identity is represented as either aip:web:<domain>/<path> or aip:key:ed25519:<multibase>, and tokens travel over MCP, A2A, and HTTP. Compact mode verification takes 0.049 ms in Rust and 0.189 ms in Python; in a real multi-agent deployment with Gemini 2.5 Flash, AIP adds 2.35 ms, or 0.086% of total end-to-end latency. In 600 adversarial attack attempts, the protocol achieved a 100% rejection rate, with delegation depth violation and audit evasion through empty context uniquely caught by the chained delegation model (Prakash, 25 Mar 2026).

The agent:// scheme solves a complementary problem: topology-independent identity and capability-based discovery. Its structure is

$\alpha = 1$5

where the trust root anchors organizational authority, the capability path encodes what the agent does, and the agent-id is a TypeID based on UUIDv7. Discovery keys are derived as

$\alpha = 1$6

yielding trust-root-scoped DHT lookup under Kademlia. The evaluation reports 100% capability coverage on 369 production tools with zero collision, discovery precision and recall of 1.0 across 10,000 agents, formal migration invariance, and all core operations under 5 microseconds (Rodriguez, 21 Jan 2026). This suggests an identity layer suitable for federated Agent Islands in which agents migrate, replicate, or move across providers without breaking references.

At the network level, “Making Cellular Networks Crisis-Proof: Towards Island-Ready, Resilient-By-Design 6G Communication Network” gives “island” an infrastructural meaning (Janzen et al., 4 Dec 2025). An island is a crisis-struck area isolated from the outside Internet but still possessing local radio and IP connectivity. The paper argues that present 5G and 5G-Advanced systems are fragile because stateful core functions and IMS are centralized; when backhaul to the central core fails, even local services such as emergency calls, messaging, and local application access become unavailable. The proposed 6G vision is island-ready: regions should be able to operate as autonomous network islands with local control plane, local IMS, local UPF, and local application servers, and to transition into and out of island operation while maintaining state. Although the paper does not explicitly discuss software agents, it directly supports an Agent Island interpretation in which edge computing and local-first application architectures host autonomous services during disconnection (Janzen et al., 4 Dec 2025).

Taken together, these works define a layered Agent Island stack: bounded execution or communication domains, topology-independent agent identities, verifiable delegation and provenance across tool and agent protocols, and resilience mechanisms that allow local operation under degraded connectivity. The recurrent research claim is not that all such systems are identical, but that boundedness—of space, trust, topology, or connectivity—can be made a first-class design primitive for multi-agent systems.