Small Agent Networks (SANs)

• SANs are distributed systems composed of small, specialized agents that make decisions based on local information and constrained communication.
• They employ diverse coordination methods—from neighbor notifications and trails to semantic orchestration across distinct layers—to balance efficiency and stability.
• Empirical studies show SANs enhance security detection and consensus dynamics while highlighting challenges in optimal parameterization and dynamic topology management.

Searching arXiv for the cited SAN-related papers and topic framing. arxiv_search: query: "Small Agent Networks SANs distributed self management distributed security systems SANNet AgentsNet small-world networks" max_results: 10 Small Agent Networks (SANs) denote distributed systems composed of many small, specialized, and largely autonomous agents that interact over an explicit communication topology and make decisions from local state, limited neighbor information, or both. Across the literature represented here, SANs appear in at least three closely related senses: as mobile artificial cells in distributed security systems, as cross-layer specialized agents coordinated for agentic networking, and as topology-constrained language-model agents studied through graph-theoretic and small-world formulations. In all cases, the defining concern is not merely multi-agent capability, but decentralized coordination under limited information, bounded communication, and heterogeneous local objectives (0805.1785, Xiao et al., 25 May 2025, Wang et al., 19 Dec 2025, Grötschla et al., 11 Jul 2025).

1. Conceptual scope and defining properties

In the distributed security formulation, the system is composed of many small entities that move through a network, each capable of detecting and mitigating one specific intrusion type. These entities work autonomously, with only local node state plus small neighbor information to guide decisions. Two recurring SAN-typical problems are explicitly identified: agents that process packets must decide when to stay versus move and, if moving, which neighbor to choose; and agents that periodically check nodes must decide which neighbor to visit next so that all nodes are checked regularly but not too frequently (0805.1785).

In the agentic networking formulation, SANs map to many specialized, small agents operating at different layers with narrow capabilities and distinct objectives. The concrete instantiation contains an application-layer agent (aAgent), a physical-layer agent (pAgent), and a network-layer agent (nAgent), all orchestrated by an agent controller that performs semantic goal detection, task separation into layer-specific subtasks, agent selection based on agent cards, subtask execution monitoring, and dynamic, weighting-based conflict resolution (Xiao et al., 25 May 2025).

In topology-centered work on LLM-based multi-agent systems, SANs are treated as networks whose communication graph is itself a first-class design variable. Here the central issue is the trade-off between local specialization and global integration: sparse local neighborhoods promote consistent subproblem solving and reduce communication cost, while a small number of shortcuts enable fast dissemination of corrections or novel evidence across clusters (Wang et al., 19 Dec 2025). AgentsNet places this perspective into a distributed-systems setting by evaluating whether LLM-based agents can agree on protocols, exchange information in rounds, and converge to globally consistent solutions on explicit graph topologies under the LOCAL model (Grötschla et al., 11 Jul 2025).

A common misconception is to treat SANs as simply “small models” or merely “many agents.” The literature instead emphasizes small, specialized agents embedded in a constrained networked decision process. Another possible confusion is terminological: “SANNet” is the name of a particular architecture—Semantic-Aware Agentic AI Networking—rather than a synonym for SANs as a general class (Xiao et al., 25 May 2025).

2. Core architectural patterns

The distributed security architecture is node-centric. The underlying network is packet-switched and company-like, with about 500 nodes. Each node has neighbors, local computational resources, and limited storage for protocol-related metadata. Numerous specialized artificial cells move across nodes; each cell can identify and act on exactly one intrusion type. Communication is limited to local node communication to resident agents, neighbor-to-neighbor forwarding of small notifications, and the absence of global centralized control. SANA instantiates this structure, combining mobile agents with common security components such as antivirus, firewalls, and IDS, while retaining decentralized movement and task decisions (0805.1785).

The SANNet architecture is controller-centric but not centralized in the classical delegation sense. Its mobile-system model focuses on three layers—application, physical, and network—each associated with an agent abstraction. The aAgent is defined by

$G^a = \langle A^a, S^a, L^a, D^a \rangle,$

with analogous tuples for $G^p$ and $G^n$. The controller detects the user’s semantic goals, performs task separation and semantic translation, queries agent cards describing function, environmental states, action space, associated layer, objective, and loss functions, selects suitable agents, dispatches subtasks, and coordinates learning and acting across layers (Xiao et al., 25 May 2025).

The topology-centric SAN formulation is graph-explicit from the outset. In the Watts–Strogatz construction, a ring of $n$ agents is formed with mean degree $k$, and each edge is rewired with probability $p$ to create shortcuts while avoiding self-loops and duplicates. This yields high local clustering and short average path lengths. The same paper maps local clusters to specialties such as math solvers, retrieval specialists, and code checkers, with shortcuts connecting epistemically distant roles so that corrections can propagate rapidly without fully dense communication (Wang et al., 19 Dec 2025).

AgentsNet formalizes the network as a graph $G=(V,E)$ with $|V|=n$, neighborhood $N(v)$, degree $\deg(v)$, maximum degree $G^p$0, and diameter $G^p$1. It uses three graph generators—Watts–Strogatz “SmallWorld,” Barabási–Albert “ScaleFree,” and Delaunay triangulations “Delaunay”—to test how topology affects collaborative reasoning, self-organization, and communication fidelity. This framing makes SANs analyzable in terms familiar from distributed systems rather than purely prompt-engineering terms (Grötschla et al., 11 Jul 2025).

Taken together, these architectures suggest a broad SAN pattern: small specialized agents, explicit locality constraints, and a coordination mechanism that is either purely local, mediated by compact metadata, or structured by graph design.

3. Local coordination and self-management mechanisms

The most explicit local self-management mechanisms appear in the distributed security work, which introduces two neighbor-information protocols.

The first protocol, notification of insufficient security, addresses the problem of approximately uniform agent distribution so that all nodes maintain minimum security coverage. Each node $G^p$2 maintains a minimum required security threshold $G^p$3 and computes

$G^p$4

If $G^p$5, the node informs resident entities about the low security level, sends a small packet to all neighbors indicating the deficient node and the numeric missing security level, and allows forwarding only while a hop-wise monotonically decreasing value remains above threshold. If multiple notifications arrive at one node in the same time step, only the highest missing-security notification is forwarded. Nodes do not store notifications. Each agent has a movement probability $G^p$6 and direction-choice function $G^p$7; upon receiving a low-security notification, $G^p$8 is increased and $G^p$9 is set toward the deficient node. In simulations, agents in a node currently below $G^n$0 do not leave that node (0805.1785).

The second protocol, trails of entities, addresses periodic node checking. Each node maintains, for each outgoing connection and entity type, a float value indicating recency and frequency of checks. After an entity of type $G^n$1 finishes a check at node $G^n$2 and moves along a link, the relevant trail value is increased via

$G^n$3

with $G^n$4 and $G^n$5, and every time step all trail values decay via

$G^n$6

with $G^n$7, until zero. Agents probabilistically prefer links with lower trail values, so recently and frequently checked neighbors become less attractive. The mechanism is purely local: nodes do not exchange trail information (0805.1785).

These two protocols instantiate a general SAN principle: minimal information can be enough if it is task-specific. The notification protocol exports only missing security level and source identity; the trail protocol exports nothing beyond local per-link state. A plausible implication is that SAN coordination does not require rich state sharing if the shared signal is tightly aligned with the local decision problem.

The same localist logic reappears in later work, although at different abstraction levels. AgentsNet’s synchronous LOCAL-style protocol gives each agent only its neighbors’ previous-round messages and requires it to generate flat JSON addressed to neighbors before producing a final structured answer after $G^n$8 rounds. Operationally, each outgoing message is a local function of prompt and received neighbor messages, which is fully consistent with SAN-style locality constraints (Grötschla et al., 11 Jul 2025).

4. Semantic orchestration and multi-objective coordination

SANNet extends the SAN idea from local movement control to semantic goal-driven cross-layer orchestration. A UE-side aAgent uses an LLM to detect user prompts, and in the prototype the agent controller uses OpenManus with the transformer-based LLM Qwen-7B to infer user QoE goals from language input such as “increase video resolution” and “make video clearer.” The controller translates the inferred semantic goal into layer-specific requirements: application-layer adjustment of client rendering parameters, physical-layer sensing of multiple channels and estimation of achievable data rates, and network-layer tracking and prediction of bandwidth between the UE and the 5GS core (Xiao et al., 25 May 2025).

The multi-agent optimization problem is formulated as

$G^n$9

where $n$0. Because objectives may conflict, SANNet seeks Pareto solutions through an MGDA reformulation:

$n$1

To quantify conflict, the framework defines a conflicting error,

$n$2

where $n$3 are the optimal MGDA weights. It then updates weights and parameters stochastically:

$n$4

$n$5

Under the stated Lipschitz assumptions, Theorem 1 gives a bound on $n$6, and setting $n$7 and $n$8 yields $n$9. SANNet also defines a generalization error $k$0 and states the bound $k$1 under the given gradient-norm condition (Xiao et al., 25 May 2025).

This formulation distinguishes SANNet from coordination schemes that only perform heuristic arbitration. Here, semantic goal inference, subtask decomposition, agent selection, and conflict resolution are integrated into a single cross-layer control loop with explicit optimization and theoretical guarantees. At the same time, the paper notes assumptions and limits: Lipschitz continuity and bounded gradient norms are required for the bounds, and the current prototype evaluates C-error and G-error rather than end-to-end QoS or QoE (Xiao et al., 25 May 2025).

A plausible implication is that SANs can be organized at more than one level of abstraction. In SANA, the coordination signal governs agent movement and task allocation directly; in SANNet, the coordinating object is a semantic goal translated into layer-specific loss interactions.

5. Communication topology as a design prior

The topology-centered literature argues that SAN performance depends not only on local policies, but on the graph over which agents exchange information. Small-world networks are defined as graphs with high local clustering and short average path lengths. For a node $k$2 with degree $k$3 and $k$4 edges among its neighbors, the clustering coefficient is

$k$5

with network-level

$k$6

If $k$7 is the shortest-path distance, the average shortest path length is

$k$8

Relative to a random graph with the same $k$9 and mean degree $p$0, small-worldness is summarized by

$p$1

The design recommendation is to remain in a regime with high $p$2, low $p$3, and $p$4 (Wang et al., 19 Dec 2025).

In multi-agent debate, the paper compares fully connected, sparse ring, random, and small-world topologies. Fully connected communication offers maximal integration but high token cost and a tendency toward error amplification and groupthink because all agents see every message. Sparse rings minimize tokens and provide local stability but have long global mixing times. Ad-hoc dynamic selection can achieve low path lengths and low token cost, and in some settings highest raw accuracy, but it produces unstable consensus trajectories. Small-world connectivity is presented as a balanced design: nearly the same accuracy and token cost as strong baselines, but significantly more stable consensus dynamics (Wang et al., 19 Dec 2025).

The same work introduces uncertainty-guided rewiring. Agents estimate semantic entropy

$p$5

derive confidence scores and belief distributions, measure pairwise epistemic divergence through Jensen–Shannon divergence and uncertainty contrast, and add shortcuts when a composite divergence exceeds threshold while preserving degree bounds, a small-worldness floor, and message budget. Consensus stability is monitored through per-round vote entropy, consensus variance, and time-to-consensus (Wang et al., 19 Dec 2025).

AgentsNet complements this design argument by making topology operational in benchmark form. It explicitly studies “SmallWorld,” “ScaleFree,” and “Delaunay” graphs and ties feasibility to graph diameter and maximum degree. Global tasks benefit from small diameter; local symmetry-breaking is sensitive to conflict resolution near high-degree nodes; Delaunay graphs ease some local coordination because of relatively low $p$6 but can slow global alignment when relying only on neighborhood exchanges (Grötschla et al., 11 Jul 2025).

This suggests that SAN design has two separable but interacting layers: local decision rules and communication topology. The 2008 security protocols optimize the former under a fixed network graph; the 2025 topology-centered work treats the graph itself as an inductive bias.

6. Empirical performance, benchmarks, and failure modes

The distributed security experiments were conducted on a packet-switching, company-like topology with about 500 nodes, with viruses, worms, and trojans arriving via network or internal attacks, and with many specialized artificial cells spanning 60 types. For notification of insufficient security, detection without notification was about 80% of intrusions identified, whereas detection with notification was about 98% of intrusions identified, and nodes achieved a guaranteed minimum security level. Compared to a centralized delegation server, notification used about 80% less bandwidth and increased detection performance by about 15% relative to the centralized approach. Proper parameterization was required to avoid overreaction and oscillation (0805.1785).

For trails of entities, only about 70% of nodes were checked without trails, with some nodes checked every second time step. With trails, all nodes were checked regularly, overly frequent checks were prevented, and the required computational power for node checks was reduced by about 60%. A key failure mode arose in sparsely connected fragments: trail values on bridge links became high, so agents avoided crossing them. Proposed remedies were increasing the decay rate on those links or disabling trails locally to give equal movement probability (0805.1785).

SANNet evaluates a hardware prototype based on open RAN and a 5GS core, with a UE using an NI 2944R USRP and desktop, a gNB using an NI 2944R USRP and workstation with an NVIDIA GeForce RTX 4090 GPU, srsRAN on UE and gNB, Open 5GS core, OpenManus with Qwen-7B for the controller, a Unity-based 3D video rendering aAgent, a multi-channel sensing/assignment pAgent, and a bandwidth tracking/prediction nAgent. In a user QoE-driven immersive communication case study, the dynamic weighting mechanism reduced C-error by up to 63% compared to static weighting as iteration count increased, while G-error increased slightly as iterations grew. Statistical significance tests were not reported (Xiao et al., 25 May 2025).

AgentsNet evaluates multi-agent reasoning on five classical tasks: $p$7-Coloring, Minimal Vertex Cover, Maximal Matching, Leader Election, and Consensus. Global tasks use $p$8 rounds, while local tasks use small fixed budgets depending on graph size. On graphs of 4–16 nodes, frontier LLMs performed strongly on global agreement, especially Consensus and often LeaderElection, while Coloring, Matching, and especially VertexCover were more challenging. For example, Gemini 2.5 Pro achieved mean AgentsNet 0.80 (0.02), with Consensus 0.99 (0.01), LeaderElection 0.89 (0.06), Matching 0.75 (0.05), Coloring 0.62 (0.07), and VertexCover 0.73 (0.06); Claude 3.7 Sonnet achieved mean AgentsNet 0.70 (0.02); GPT-4.1 mini achieved mean AgentsNet 0.45 (0.01). Extending Gemini 2.0 Flash to 20–100 agents showed monotonic decline across all tasks, with near-zero success at $p$9 (Grötschla et al., 11 Jul 2025).

Failure analyses across these works are notably consistent. AgentsNet reports protocol misalignment or late strategy agreement, over-trusting neighbor information, stale state assumptions, synchronous message lag, and output formatting errors (Grötschla et al., 11 Jul 2025). The security paper reports oscillation from overly aggressive movement responses and bottleneck starvation induced by trails (0805.1785). The small-world work identifies over-shortcutting, miscalibrated uncertainty signals, and task regimes in which small-world priors may underperform denser topologies (Wang et al., 19 Dec 2025). These are not identical mechanisms, but they all point to the same SAN tension: insufficient coordination produces fragmentation, while excessive coupling produces instability or unnecessary overhead.

7. Design principles, limitations, and open directions

Several practical design guidelines recur across the literature. In distributed security SANs, minimal neighbor information is sufficient if it directly encodes what agents need to decide; movement probability should increase in response to deficit signals but be capped to prevent floods and oscillations; when multiple deficiency notifications arrive, only the highest-need one should be forwarded; trails should remain local-only, increasing exponentially on use and decaying linearly over time; bottleneck starvation should be mitigated by parameter changes or disabling trails locally; and notifications should remain stateless so that decisions rely on fresh signals (0805.1785).

In SANNet, agents should remain small and specialized per layer, with clear act spaces, state spaces, and explicit loss functions; agent cards should expose capabilities, objectives, and environmental constraints; semantic goals should be mapped to a minimal set of layer-specific subtasks; and dynamic weighting should be used to resolve conflicts, with step sizes tuned as $G=(V,E)$0 and $G=(V,E)$1 to guarantee the stated convergence rate (Xiao et al., 25 May 2025).

In topology-aware SANs, small-world connectivity should be treated as a first-class prior: begin with a Watts–Strogatz graph, choose $G=(V,E)$2 and $G=(V,E)$3 according to task difficulty, bridge heterogeneous roles with a limited number of shortcuts, and monitor $G=(V,E)$4, consensus variance, and token cost to remain in the small-world regime. Shortcut edges should preferentially carry synthesized summaries or uncertainty-resolving queries rather than unrestricted raw text in order to control token cost (Wang et al., 19 Dec 2025).

AgentsNet adds prompt- and protocol-level guidance: round budgets for global agreement tasks should scale approximately with $G=(V,E)$5; explicit symmetry-breaking rules in the initial prompt reduce late negotiation; strict JSON schemas and final-answer templates reduce non-semantic failures; and message hygiene should encourage agents to state current commitments, broadcast changes, and challenge inconsistent neighbor claims (Grötschla et al., 11 Jul 2025).

The limitations are equally clear. The security work does not provide statistical confidence, variance, or explicit formulas for $G=(V,E)$6 and $G=(V,E)$7, and does not specify the exact decrement constants for the notification protocol (0805.1785). SANNet assumes Lipschitz continuity and bounded gradient norms for its theoretical guarantees and does not yet report end-to-end QoS or QoE statistics (Xiao et al., 25 May 2025). The small-world study calls for formal bounds linking $G=(V,E)$8, $G=(V,E)$9, and $|V|=n$0 to consensus stability and accuracy in stochastic debate dynamics (Wang et al., 19 Dec 2025). AgentsNet simplifies coordination to a synchronous LOCAL protocol, assumes cooperative non-adversarial agents, and leaves asynchronous communication, heterogeneity, Byzantine resilience, and dynamic topology as open directions (Grötschla et al., 11 Jul 2025).

Taken together, these works define SANs not as a single architecture but as a design regime. The regime is characterized by many small specialized agents, explicit topological constraints, local or semantically compressed coordination signals, and a persistent attempt to balance specialization, integration, bandwidth, and stability. This suggests that SAN research is converging on a common question: how much structure—informational, semantic, or topological—is necessary for decentralized agents to behave as a coherent system without sacrificing the robustness and efficiency that decentralization is meant to provide.