AgentHub: Multi-Agent Infrastructure

• AgentHub is a layered infrastructure that enables both registry-based discovery and execution coordination for software agents.
• It integrates metadata, lifecycle controls, and evidence-linked discovery to support governance, trust, and interoperability.
• AgentHub’s design spans decentralized discovery, execution orchestration, and domain-specific adaptations, driving efficient multi-agent workflows.

Searching arXiv for papers on AgentHub and closely related agent-hub infrastructure. AgentHub denotes a family of infrastructures for publishing, discovering, evaluating, governing, composing, and operationalizing software agents. In its most explicit formulation, it is proposed as “a research agenda for agent sharing” and as a registry intended to make agents “shareable, trustworthy, and composable like software libraries” (Pautsch et al., 3 Oct 2025). Contemporaneous work uses the same term, or closely related “Agent Hubs,” for a globally distributed registry and discovery substrate in the AIOS Internet of AgentSites (Zhang et al., 19 Apr 2025), a telecom-specialized low-code multi-agent platform built around TeleMCP (Shah et al., 12 Nov 2025), an execution-focused coordination unit inside the Agent Exchange economy (Yang et al., 5 Jul 2025), a trustworthy mixed-initiative hub blueprint grounded in HAX (Scibelli et al., 12 Dec 2025), and an orchestration-centric design that maps instruction, context, tools, and model selection into a reusable hub kernel (Ruan et al., 3 Feb 2026). Taken together, the literature presents AgentHub as a design space spanning registry-layer metadata, decentralized discovery, workflow orchestration, governance, and economic coordination.

1. Definitions and major usages

The term is not used in a single, uniform sense. Some papers treat AgentHub primarily as registry infrastructure; others use it to denote an execution and coordination layer. The resulting landscape is best understood through the specific roles assigned to the term in recent work.

Source	Meaning of AgentHub	Primary emphasis
Research agenda (Pautsch et al., 3 Oct 2025)	Registry to support production and consumption of software agents	Capability metadata, lifecycle, governance
AIOS-IoA (Zhang et al., 19 Apr 2025)	Globally distributed registry, index, and discovery substrate	DHT/Gossip discovery, liveness, endpoints
Tele-LLM-Hub (Shah et al., 12 Nov 2025)	Domain-specialized agent hub for telecom	TeleMCP, low-code orchestration, tool integration
Agent Exchange (Yang et al., 5 Jul 2025)	Operational coordinators competing in AEX-hosted auctions	Team formation, bidding, intra-hub allocation
HAX (Scibelli et al., 12 Dec 2025)	Blueprint for a hub/platform treating agents as colleagues	Trust, transparency, mixed initiative
AOrchestra (Ruan et al., 3 Feb 2026)	Mapping from orchestration tuple to AgentHub components	Dynamic sub-agent creation, cost-aware routing

This distribution of meanings suggests that AgentHub is less a single product category than a layered architectural pattern. At minimum, the pattern addresses agent discovery and capability description; in stronger formulations it also includes evidence, lifecycle controls, policy enforcement, execution interfaces, and multi-agent coordination.

2. Metadata-first registries and discovery planes

The registry-centric formulation places metadata at the center of agent sharing. The proposed AgentHub research agenda argues that existing model hubs “expose artifacts and informal model cards,” but “omit normalized dependency and capability schemas, hampering reuse.” It therefore advocates a “metadata-first registry” with a canonical manifest covering capabilities and atomic actions, input–output modalities, protocol roles and bindings, runtime permissions, environment bindings, and SBOM-style dependencies on models, tools, datasets, services, and other agents. The same agenda adds persistent identifiers, signed manifests, public transparency logs, re-executable evidence pipelines, lifecycle states such as active, deprecated, rotated, retired, and revoked, and APIs for publication, update, deprecation or revocation, and evidence submission (Pautsch et al., 3 Oct 2025).

A concrete discovery plane appears in AIOS-IoA. There, AgentHub is the registry and discovery layer that stores and disseminates agent metadata including agent_id, description or capabilities, last_seen, node_id, node_ip, and node_port, while supporting optional extensions such as mcp_tools, rpc_endpoint, model, version, trust/reputation score, geographic hints, policy metadata, and schema version. The interface is expressed through JSON-RPC 2.0 methods including agenthub/registerAgent, agenthub/updateAgent, agenthub/findAgents, and agenthub/getAgentById, and the protocol basis combines MCP v1.2.1 with JSON-RPC 2.0 (Zhang et al., 19 Apr 2025).

A common point across these formulations is that discovery is not reduced to name lookup. Capability descriptors, evidence, lifecycle state, provenance, and endpoint metadata are treated as first-class registry objects. This distinguishes AgentHub from narrower naming or negotiation layers.

3. Distributed execution and orchestration

In the AIOS implementation, AgentHub is coupled tightly to execution hosts. AIOS Server is the runtime that hosts agents and communicates via MCP plus JSON-RPC; AgentHub runs alongside these server instances, and all AgentHub nodes are essentially AIOS Server nodes with registry functionality activated. Discovery is supported by a Kademlia-like DHT with XOR distance, k-buckets, replication, and expected lookup hops of $O(\log N)$, plus a Gossip protocol for near real-time presence synchronization, eventual consistency, and resilience against single points of failure (Zhang et al., 19 Apr 2025).

A different execution-centered formulation appears in AOrchestra, which models any agent as the four-tuple

$\Phi = (I, C, T, M),$

where $I$ is instruction, $C$ is curated context, $T$ is the tool set, and $M$ is the underlying model. The orchestrator itself operates over the structured action space

$$\mathcal{A}_{\text{AOrchestra}} = \{\mathrm{Delegate}(\Phi),\ \mathrm{Finish}(y)\},$$

spawning tailored executors on demand rather than relying on fixed sub-agent roles. The paper maps this directly to an AgentHub design through instruction templates, a context store, a tool registry, a model registry, and an executor spawner, with explicit attention to performance–cost trade-offs and plug-and-play execution backends (Ruan et al., 3 Feb 2026).

HAX adds a human-facing orchestration layer. Its blueprint for building AgentHub is organized into three phases: Design Heuristics, Agentic SDK, and Behavioral Proxy. The heuristics—Control, Clarity, Recovery, Collaboration, and Traceability—are paired with a schema-driven SDK that enforces structured outputs via tool-call contracts, and with a proposed behavioral proxy that schedules, filters, and contextualizes agent activity. Grounding in TIP theory organizes visibility around Inception, Problem Solving, Conflict Resolution, and Execution, thereby making user approval, trade-off inspection, trust repair, and rollback native parts of hub behavior rather than afterthoughts (Scibelli et al., 12 Dec 2025).

These works collectively move AgentHub beyond static cataloging. The hub becomes the place where discovery, delegated execution, state curation, and human oversight are integrated.

4. Governance, security, and lifecycle transparency

Governance and security are central rather than peripheral concerns in the registry-oriented literature. The AgentHub research agenda imports lessons from package registries and supply-chain standards, emphasizing signed metadata, verified namespaces, auditable publication and revocation decisions, fast revocation and key rotation across mirrors, The Update Framework (TUF), and structured provenance attestations such as “SLSA v1.1 Verification Summary Attestation (VSA).” Its threat model includes typosquatting, account hijacks, malicious updates, package confusion, prompt injection, “multi-agent prompt infection,” data exfiltration, privacy leaks, tool misuse, privilege escalation, unauthorized code execution, and deserialization exploits (Pautsch et al., 3 Oct 2025).

The AIOS prototype addresses a narrower operational slice. It currently emphasizes schema validation and multi-registry observation, with future work explicitly listing cryptographic identities, signed registrations, verification by registry nodes, and OAuth2/JWT for centralized registry deployments. It also proposes rate limiting per node_id and IP, reputation or trust fields, blacklisting and whitelisting, and community moderation, while exposing only minimal metadata by default and supporting delisting or opt-out behavior (Zhang et al., 19 Apr 2025).

Tele-LLM-Hub operationalizes governance through context minimization. TeleMCP nodes normalize raw telecom artifacts into canonical typed objects and let users select which fields to publish rather than “blasting raw blobs.” All objects carry provenance tags, Output nodes capture complete traces, and Conditional nodes such as Human Approval gate critical steps before automation proceeds. In this design, provenance and field-level publication control are the mechanisms through which privacy, auditability, and governance enter multi-agent execution (Shah et al., 12 Nov 2025).

HAX complements these infrastructure controls with interface-level trust mechanisms. Intent preview, permission gates and autonomy knobs, inline rationales and uncertainty displays, trust repair triggers, provenance timelines, visual diffs, and progressive disclosure are presented as ways to make agent behavior understandable, recoverable, and alignable with human goals (Scibelli et al., 12 Dec 2025).

5. Domain-specific and economic forms

Tele-LLM-Hub exemplifies a domain-specialized AgentHub. It couples TeleMCP, the Telecom Model Context Protocol, with a low-code canvas that unifies agent creation, multi-agent orchestration, and tool integration for 5G and O-RAN stacks such as srsRAN and OpenAirInterface. Its architecture includes a direct chat interface, a repository of pre-built systems, an Agent Maker using RANSTRUCT-based fine-tuning, and an MA-Maker with Input, Agent, TeleMCP, Logic, Conditional, and Output nodes. TeleMCP standardizes telecom artifacts into typed canonical objects such as procedural-flow, log-window, and message-record, each with common provenance metadata including source, transform, published fields, version, and creation time. This hub design targets specification-aware testing and validation, anomaly detection, resource management, slice control, and security across telecom artifacts such as KPIs, logs, traces, PCAPs, and telemetry (Shah et al., 12 Nov 2025).

In the Agent Exchange ecosystem, by contrast, Agent Hubs are not registries but operational coordinators. They consume structured task specifications from the User-Side Platform, capability and performance signals from the Agent-Side Platform, and shared knowledge and attribution infrastructure from the Data Management Platform, while competing in AEX-hosted multi-attribute auctions. Coordination is organized as a two-stage process: first, multiple hubs compete to deliver complete task solutions; second, the winning hub performs fine-grained internal assignment and collaboration. The paper proposes combinatorial optimization for internal team formation, multi-attribute agent auctions beyond price, and Shapley value computation for internal credit settlement. It also formalizes cost feasibility and cooperative surplus through constraints such as

$\sum_{i \in A} \mathrm{Cost}_i \leq V(A, T)$

and

$V(A, T) > \sum_{i \in A} V(\{i\}, T).$

Here, AgentHub denotes a coordination and allocation mechanism embedded in an agent marketplace rather than a discovery catalog (Yang et al., 5 Jul 2025).

These domain-specific and economic variants show that the hub abstraction is extensible. One variant specializes context protocols and low-code workflows for telecom; another specializes mechanism design, bidding, and attribution for agent economies.

6. Evaluation, limitations, and open questions

Reported evaluations are heterogeneous because the papers target different layers of the stack. The AIOS deployment reports MCP/JSON-RPC latency under 200 ms across loads, throughput up to 229 req/s in cloud deployment, and decentralized registration latency of roughly 1 ms average across 3, 5, and 7 nodes. The AEX simulation study covers 1,350 trials and reports significant differences among methods, with ANOVA $F(4,2245)=12.7$, $\Phi = (I, C, T, M),$0; its Enhanced Auction achieves quality $\Phi = (I, C, T, M),$1, cost efficiency $\Phi = (I, C, T, M),$2, robustness $\Phi = (I, C, T, M),$3, and adaptability High. AOrchestra, which explicitly maps its orchestration kernel to an AgentHub design, reports that training-free use with Gemini-3-Flash achieves a 16.28% relative improvement against the strongest baseline across GAIA, Terminal-Bench, and SWE-Bench-Verified, with Avg Pass@1 $\Phi = (I, C, T, M),$4, and that its ICL setting reduces average cost by 18.5% while improving accuracy by $\Phi = (I, C, T, M),$5 points. Tele-LLM-Hub, by contrast, does not report quantitative latency, throughput, or reliability metrics for the platform itself, emphasizing reproducibility, auditability, and domain grounding instead (Zhang et al., 19 Apr 2025, Yang et al., 5 Jul 2025, Ruan et al., 3 Feb 2026, Shah et al., 12 Nov 2025).

Several limitations recur across the literature. The registry agenda is explicitly motivated by fragmentation and by the absence of normalized capability schemas, lifecycle transparency, and evidence-linked discovery. The AIOS prototype lacks full cryptographic identity and signed registrations. TeleMCP standardization is described as ongoing, and broader stack and hardware support remains future work. AEX evaluates under static capabilities, perfect information, honest reporting, and stabilized market conditions rather than production assumptions. HAX states that formal evaluations of trust patterns and benchmarking of agentic interactions remain future work, and its behavioral proxy is proposed but left beyond the scope of the paper (Pautsch et al., 3 Oct 2025, Zhang et al., 19 Apr 2025, Shah et al., 12 Nov 2025, Yang et al., 5 Jul 2025, Scibelli et al., 12 Dec 2025).

A persistent misconception is to equate AgentHub with a protocol directory alone. The registry agenda explicitly distinguishes its scope from MCP Registry, ANS, ACNBP, and NANDA, arguing that naming and negotiation do not by themselves provide lifecycle governance, structured manifests, signed provenance, evidence pipelines, or workflow integration. Another misconception is to assume centralization is intrinsic: the AIOS system shows a decentralized discovery substrate built on DHT and Gossip, while the governance proposals in the registry literature discuss federation and verified namespace control. A third misconception is to reduce AgentHub to orchestration only: the economic and telecom variants show that internal assignment, attribution, field-level publication control, auditability, and domain semantics can be constitutive parts of the hub abstraction rather than optional add-ons (Pautsch et al., 3 Oct 2025).

The literature therefore converges less on a single implementation than on a set of recurring requirements: capability clarity, lifecycle transparency, interoperability, governance, security, and workflow integration. This suggests that future AgentHub systems will be judged not merely by whether they can locate or invoke agents, but by whether they can make agent capabilities legible, evidence-linked, policy-constrained, revocable, composable across runtimes, and auditable across long-lived multi-agent workflows.