Source-linked AI summary
Internet of Agents: Fundamentals, Applications, and Challenges
Yuntao Wang, Shaolong Guo, Yanghe Pan, Zhou Su, Fahao Chen, Tom H. Luan, Peng Li, Jiawen Kang, Dusit Niyato
TL;DR
The paper addresses the need for unified, agent-centric infrastructure as autonomous agents proliferate across virtual and physical environments. It surveys the Internet of Agents through its architecture and operational enablers, then identifies challenges and research directions for resilient and trustworthy ecosystems.
Problem
The proliferation of heterogeneous autonomous agents creates a need for infrastructure supporting seamless interconnection, discovery, coordination, and collaboration at scale.
Method
The survey presents a hierarchical IoA architecture and examines capability discovery, task orchestration, adaptive communication, consensus, incentives, and regulation.
Results
The paper synthesizes IoA foundations, applications, enabling technologies, and open challenges across scalability, interoperability, security, privacy, economics, and ethics.
Takeaways & Limitations
IoA development requires standardized interoperability, adaptive coordination, consensus mechanisms, and combined technical and legal safeguards for trustworthy agent collaboration.
Abstract
from arXiv · showhide
With the rapid proliferation of large language models and vision-language models, AI agents have evolved from isolated, task-specific systems into autonomous, interactive entities capable of perceiving, reasoning, and acting without human intervention. As these agents proliferate across virtual and physical environments, from virtual assistants to embodied robots, the need for a unified, agent-centric infrastructure becomes paramount. In this survey, we introduce the Internet of Agents (IoA) as a foundational framework that enables seamless interconnection, dynamic discovery, and collaborative orchestration among heterogeneous agents at scale. We begin by presenting a general IoA architecture, highlighting its hierarchical organization, distinguishing features relative to the traditional Internet, and emerging applications. Next, we analyze the key operational enablers of IoA, including capability notification and discovery, adaptive communication protocols, dynamic task matching, consensus and conflict-resolution mechanisms, and incentive models. Finally, we identify open research directions toward building resilient and trustworthy IoA ecosystems.
I. INTRODUCTION
The survey positions the Internet of Agents (IoA) as agent-centric infrastructure for interconnecting, discovering, and coordinating heterogeneous agents at scale. It presents IoA architecture and operational enablers while identifying challenges and open directions for large-scale deployment.
- Motivation: IoA connects geographically distributed, heterogeneous agents and supports autonomous discovery, task orchestration, and collaborative reasoning beyond isolated multi-agent simulations.The framework targets agents with diverse compute, network, sensing, and energy profiles across virtual and physical environments.
- Motivation: Agent-native interfaces replace human GUI emulation with APIs or semantic protocols for direct interaction among agents and Internet resources.This approach addresses the screen-scraping overhead associated with click- and keystroke-based computer-use agents.
- Research Gap: The survey addresses ecosystem isolation, single-device simulation, and rigid communication and coordination in existing multi-agent-system research.These limitations restrict third-party integration, geographic distribution, and adaptation to dynamic agent states.
- Contributions: The paper provides a hierarchical IoA architecture, distinguishes it from the traditional Internet, and surveys its characteristics and emerging applications.Its architecture overview is intended to explain how diverse, distributed agents can be integrated.
- Contributions: The survey analyzes capability discovery, dynamic task orchestration and matching, adaptive communication, consensus and conflict resolution, economic models, and regulation models.These mechanisms are presented as critical enablers for IoA operation.
- Contributions: The paper identifies unresolved IoA issues and future research directions needed to support broad adoption and sustainable evolution.The open-challenges contribution complements the survey’s architecture and enabling-technology analyses.
B. Paper Organization
The paper surveys IoA architecture, agent types and functional modules, then organizes its discussion around enabling technologies, comparisons, and applications. Its architecture connects virtual and embodied agents through layered infrastructure and gateway-mediated coordination.
- Architecture and scope: IoA connects autonomous virtual and embodied agents across diverse domains for communication, discovery, coordination, and task orchestration.
- Agent types: Virtual agents operate digitally, whereas embodied agents perceive and act in the physical world through sensors and actuators.
- Functional modules: Each agent typically integrates planning, memory, interaction, and action modules for reasoning, adaptation, communication, and execution.
- Hierarchical architecture: IoA is organized hierarchically around interconnected sub-IoAs, gateway proxy agents, and four layers: infrastructure, management, coordination, and applications.
- Gateway functions: Gateway proxy agents provide resource registration, discovery, semantic matching, domain coordination, cross-domain relays, and state synchronization.
B. Key Characteristics of IoA
IoA is characterized by autonomous, dynamic, heterogeneous, scalable, semantic-aware, and task-driven collaboration. Agents advertise capabilities, form adaptive teams, communicate through contextual representations, and reconfigure workflows as conditions change.
- Autonomous intelligence: Agents can advertise capabilities, discover complementary partners, negotiate protocols, resolve conflicts, and reach consensus without human intervention.
- High dynamics: IoA agents dynamically create, migrate, terminate, and reorganize teams, capabilities, roles, and communication patterns in response to changing tasks and resources.
- High heterogeneity: Heterogeneous agents differ in compute, energy, sensing, modalities, and communication links, requiring adaptive workload assignment, modality alignment, and protocol negotiation.
- Large-scale scalability: Scalability relies on hierarchical sharding, elastic resource orchestration, task-oriented overlays, and fault-tolerant workload rerouting.
- Semantic-aware communication: Semantic-aware communication exchanges contextual, task-relevant information and uses persistent memory to compress messages and preserve shared context.
- Task-driven cooperation: Task-driven cooperation matches agents to tasks using semantic metadata, workload, proximity, reliability, and QoS, while supporting autonomous decomposition, replanning, and topology optimization.
C. Key Differences Among Internet, IoT, and IoA
IoA differs from the traditional Internet and IoT in its objectives, architecture, discovery mechanisms, and interaction model. It emphasizes autonomous, semantic, task-driven collaboration among dynamic agent networks rather than human-centric access or device automation alone.
- Core objectives: The traditional Internet prioritizes host and information connectivity, while IoT focuses on device monitoring and control; IoA emphasizes agent collaboration and knowledge exchange.
- Architecture: IoA uses hybrid decentralization with P2P overlays and proxy gateways, allowing agents to join, leave, announce capabilities, and collaborate across domains on demand.
- Addressing and discovery: Traditional Internet discovery uses static addresses and URLs, IoT adds automated device discovery, and IoA uses dynamic identifiers, capability labels, semantic indexes, and task-driven matching.
- Interaction modes: Internet interactions are human-driven, IoT interactions are event-triggered, and IoA interactions are proactive, goal-oriented, semantic, and autonomously negotiated.
- Conceptual distinction: IoA extends beyond MAS, the Agentic Web, and Web 3.0 by combining autonomy, adaptivity, and economic incentives in an Internet-like collaboration fabric.
1) Typical Multi-Agent Frameworks:
The paper reviews representative academic and industrial IoA implementations, illustrating layered registries, orchestration engines, gateway services, and domain applications. These examples span agent communication within subnets, external coordination, and cross-domain networking.
- Framework landscape: Existing agent frameworks support autonomous reasoning, conversational multi-agent interaction, and tool integration, but IoA implementations extend these ideas toward distributed cross-domain coordination.
- Academic implementations: The IoA prototype combines agent registration and semantic matching with autonomous team formation, subtask allocation, workflow sequencing, and context-aware data access.
- Academic implementations: DAWN uses Principal and Gateway Services Agents, with gateways maintaining registries and exposing distributed tools, agents, and applications through RESTful endpoints.
- Industrial implementations: Industrial examples apply multi-agent coordination to production design, engineering, maintenance, scheduling, energy management, quality assurance, and operations.
- Application scenarios: IoA applications include smart homes, healthcare, factories, smart grids, and cities, covering subnet communication, external-agent coordination, and cross-domain networking.
- Interaction lifecycle: Cross-domain interaction uses registration, capability notification and discovery, gateway-assisted interaction, state synchronization, and protocol or consensus mechanisms.
III. WORKING PARADIGMS OF INTERNET OF AGENTS
IoA capability discovery combines evaluation, notification, and retrieval to support effective task assignment and collaboration among dynamic agents. The main trade-off is balancing reliability, freshness, privacy, communication overhead, and scalability.
- Capability Evaluation: Capability evaluation constructs agent profiles through self-reported declarations and systematic verification.Self-reporting enables rapid onboarding but may be inaccurate; standardized verification improves reliability while introducing scalability and privacy concerns.
- Capability Notification: Capability notification keeps agent capability information synchronized as agents acquire tools, skills, or model upgrades.Proactive, event-triggered, and periodic strategies provide different balances between freshness and communication efficiency.
- Capability Notification: Event-triggered updates reduce unnecessary communication overhead by updating profiles after significant events rather than continuously.Periodic consistency checks can complement event-triggered updates by maintaining a reliable global capability view.
- Capability Notification: Rapidly evolving capabilities make real-time synchronization difficult because IoA must balance freshness, scalability, and efficiency.Frequent updates can impose substantial communication overhead in large-scale systems.
3) Capability Retrieval of Agents:
Capability retrieval selects agents whose capabilities match task requirements, while interaction structure determines how selected agents collaborate. IoA combines retrieval strategies with interaction modes and network topologies suited to task characteristics.
- Capability Retrieval: Agent capability retrieval includes traditional, semantic, knowledge-based, and agentic-enhanced strategies for heterogeneous task requirements.These approaches progress from profile matching toward semantic understanding, knowledge integration, and adaptive retrieval with planning and tool use.
- Capability Retrieval: Capability discovery is a prerequisite for effective task allocation and collaborative orchestration, with retrieval improving agent-selection accuracy and flexibility.Evaluation and notification support the profiles that retrieval uses, while frequent updates create communication and consistency challenges.
- Interaction Structure Design: IoA interaction modes include aggregate, reflect, debate, and tool-use, each organizing information exchange or external-resource access differently.Aggregation combines independent predictions; reflection iteratively incorporates feedback; debate challenges predictions; tool-use invokes resources such as RAG or code executors.
- Interaction Structure Design: Communication topologies include chain, star, tree, and graph, ranging from linear or centralized control to hierarchical and peer-to-peer coordination.Graph structures support arbitrary connections and dynamic coordination, while chain, star, and tree provide progressively different control arrangements.
- Interaction Structure Optimization: Interaction modes and communication topologies should be optimized according to task characteristics because dense connectivity can create communication overhead and scalability limits.Pruning methods use gating, top-K scheduling, communication-cost penalties, or separate communication channels to reduce unnecessary interactions.
- Interaction Mode Optimization: Multi-agent debate has been applied through asymmetric debater-judge and symmetric equal-status structures for collective decision-making.The surveyed applications include machine translation and mathematical reasoning, while empirical studies report more truthful and accurate conclusions than single-agent reasoning.
2) Task Orchestration:
IoA task orchestration decomposes high-level requests into executable subtasks and allocates them using agent capabilities, context, and resource availability. Approaches range from rule-based and routing-based control to learning-based and self-organizing coordination.
- Task Orchestration: Task orchestration consists of task decomposition and task allocation, linking high-level requests to capability- and context-aware agent assignments.Resource availability is also considered when allocating subtasks.
- Task Decomposition: Rule-based decomposition applies predefined schemas or logical structures to provide fine-grained control over complex objectives.Examples combine task parsing, virtual sub-agent instantiation, semantic query rewriting, and schema-guided augmentation, but can struggle in changing environments.
- Task Decomposition: Learning-based decomposition infers decomposition strategies through environmental interaction and can jointly learn decomposition and sub-agent policies.This approach improves sample efficiency and removes the need for manually designed rules.
- Task Allocation: Task allocation uses routing-based strategies to select suitable agents or models, offering efficiency and scalability but depending on routing quality.RouteLLM, Hybrid-LLM, and RouterEval represent model-selection routing and routing evaluation.
- Task Allocation: Self-organizing allocation lets agents coordinate resource use and task distribution without centralized control, but introduces coordination overhead and convergence instability.Examples optimize cost-efficiency or communication through artificial-economy and reinforcement-learning paradigms.
C. Communication Protocols for IoA
IoA communication protocols address the isolation of existing agent frameworks by standardizing message exchange, negotiation, context sharing, identity management, and external-tool integration. MCP and A2A are representative protocols, with MCP using a client-server workflow for capability discovery and tool or resource access.
- Communication Protocols: Existing agent frameworks are often ecosystem-isolated and distributed across heterogeneous devices and contexts, limiting interoperability and third-party integration.Standardized protocols are proposed as the backbone for scalable and interoperable IoA ecosystems.
- Representative Protocols: Representative IoA communication protocols include MCP, A2A, and ANP, which support cross-agent message exchange and coordination.The survey presents MCP and A2A through their respective workflows and identifies protocol standardization as central to IoA interoperability.
- Anthropic’s MCP: MCP is an open standard for integrating LLMs with external tools, data sources, and systems through a unified interface.OAuth-based authorization enables access to real-time data and services without exposing user credentials.
- Anthropic’s MCP: MCP uses a client-server architecture in which a host client discovers server capabilities and requests tools or resources as needed.Servers expose executable tools, data resources, or predefined prompts; they execute requested functions or retrieve resources and return results.
2) Google’s A2A:
IoA communication protocols support interoperability, agent discovery, and collaboration among heterogeneous agents, while consensus mechanisms regulate interaction, align reasoning, and scale coordination.
- Google’s A2A:: A2A provides a networking layer for heterogeneous-agent interoperability, direct communication, knowledge sharing, and authenticated collaboration.It differs from MCP, which connects individual LLMs with external data sources and tools.
- Google’s A2A:: A2A collaboration begins when a client agent retrieves a remote agent’s standardized agent card containing capabilities, skills, endpoint, and authentication metadata.The client then transmits tasks for execution or relevant information.
- Google’s A2A:: IoA protocols pursue interoperable and trustworthy ecosystems through distinct designs, including decentralized networking, standardized invocation, and adaptive communication routines.The surveyed protocols share the goal of extensible collaboration despite architectural differences.
- Consensus & Conflict Resolution: Consensus requires communication regulation and reasoning alignment to manage turn-taking, conflicts, inconsistent outputs, hallucinations, and large-scale coordination.Polling and arbitration regulate interaction, while self-consistency and collective reasoning improve reliability.
- Consensus & Conflict Resolution: Hierarchical consensus improves scalability by forming local cluster agreements and aggregating results through higher-level coordinators.This reduces communication overhead while maintaining coherent system-wide decisions.
E. Economic Models of IoA
IoA economic models align agent incentives with collective objectives through pricing, rewards, penalties, and reputation mechanisms, while contribution estimation remains difficult at scale.
- Economic Models of IoA: Economic models use dynamic pricing, incentives, and penalties to encourage honest participation and suppress selfish or inactive behavior.These mechanisms are intended to sustain long-term collaboration in decentralized ecosystems.
- Economic Models of IoA: Capability-based pricing charges for resource consumption such as token usage, interactions, and model complexity.The pricing basis reflects the computational load imposed by each agent.
- Economic Models of IoA: Contribution-aware pricing rewards agents according to marginal impact on task outcomes, often estimated with Shapley values.Agents contributing more to critical task components receive larger shares of collective rewards.
- Economic Models of IoA: IoA incentive designs draw on auctions, contracts, games, reputation, and token-based rewards to allocate resources and align agent behavior.These approaches address competition, information asymmetry, coalition benefits, historical performance, and ecosystem sustainability.
- Economic Models of IoA: Penalties can guide decentralized agents toward global objectives, mitigate conflicts, enforce feasibility, and deter free-riding through optimization or blockchain mechanisms.Blockchain-based slashing can automatically apply token forfeiture or reputation reduction.
F. Trustworthy Regulation in IoA
Trustworthy IoA regulation combines technical identity, verification, monitoring, and auditing with legal and ethical frameworks for accountability across autonomous agents.
- Trustworthy Regulation in IoA: Autonomous agents’ ability to act in digital and physical environments creates requirements for identity verification, behavior governance, and protection against malicious actions.The absence of legal accountability makes trustworthy regulation particularly important.
- Trustworthy Regulation in IoA: Accountability requires lifecycle traceability through agent identifiers, real-time monitoring, activity logs, and allocation of responsibility among providers and deployers.These mechanisms support post-incident forensics and address the many-hands problem.
- Trustworthy Regulation in IoA: DIDs, verifiable credentials, and blockchain support self-sovereign identity, privacy-aware credential exchange, immutable action records, and transparent auditing.Smart contracts can enforce authorization, behavior verification, accountability, and governance policies.
- Trustworthy Regulation in IoA: Legal frameworks should define ownership, liability, oversight, certification, regulatory sandboxes, dynamic rules, audits, and explainability requirements.Cross-border deployment remains complicated by jurisdictional fragmentation and multi-agent delegation.
- Standardization and Interoperability: Unified standards remain necessary for cross-platform discovery, capability negotiation, service composition, authentication, and interoperability.Current fragmentation and proprietary interfaces can create vendor lock-in and hinder trust across domains.
- Secure & Adaptive Communication: Adaptive communication protocols face a versatility-efficiency-portability trade-off alongside tool-poisoning, Sybil, and hallucination vulnerabilities.The survey calls for rigorous security guarantees in future agent communication.
C. Decentralized and Self-Governing Agent Ecosystems
Decentralized IoA ecosystems aim to self-organize through consensus, adaptive infrastructure, incentives, and security mechanisms, but face coordination, privacy, cyber-physical, and accountability challenges.
- Decentralized and Self-Governing Agent Ecosystems: Future decentralized IoA ecosystems may combine decentralized consensus, adaptive blockchain sharding, and bio-inspired swarm intelligence for self-organization.The target settings include smart cities, supply chains, and energy grids operating without central authority.
- Decentralized and Self-Governing Agent Ecosystems: Full decentralization improves robustness but can hinder real-time coordination and produce conflicts between local priorities and global stability.The survey highlights delivery-speed versus pedestrian-safety priorities as an example.
- Agent Economics: Agent-native economies should combine short-term capacity-aware rewards with long-term reputation tracking to foster trust and deter malicious participation.Decentralized markets must also address Sybil attacks, collusion, resource monopolization, and cross-currency interoperability.
- Privacy-Preserving Agent Interactions: Privacy-preserving IoA requires context-aware, task-driven protection because raw data, knowledge, and inference sharing can expose sensitive information.Secure multiparty computation offers strong guarantees but may introduce delays incompatible with real-time applications.
- Cyber-Physical Secure IoA: Cyber-physical IoA defenses must connect network indicators with physical consequences because compromised agents can trigger cascading failures.Legacy industrial hardware also requires retrofittable, low-overhead trust anchors.
- Ethical and Responsible Agents: Ethical IoA deployment requires transparent moral reasoning, semantic alignment, tamper-proof decision provenance, and frameworks assigning responsibility for collective harm.These challenges are especially acute in mission-critical domains and heterogeneous agent chains.
- Conclusions: The survey identifies scalability, interoperability, economics, security, privacy, and ethics as core open challenges for IoA research.It calls for continued innovation in networking architectures, interoperability standards, and security paradigms.