ETD

• Agent Discovery
• Agentic AI
• Behavioural Reputation
• Blockchain
• Cybersecurity
• Distributed Ledger Technology
• Task Delegation
• Trust Infrastructure

Throughout much of its history, the field of Artificial Intelligence (AI) has been a dynamic and fast-growing topic, improving at the same rate as technological advancements across different eras. In recent years, thanks to great progress in Large Language Models (LLMs), AI has been able to go through a foundational paradigm shift from stateless, reactive, and bounded assistants towards autonomous agents that can reason, plan, and act on their own. Coupled with LLMs as their working brain, these assistants, better known as AI agents, integrate advanced reasoning, long-term memory, tool use, and dynamic planning capabilities that have decisively broken the paradigm and definitely amplified previous AI systems. Said agents are also capable of decomposing high-level, complex objectives into granular sub-tasks, executing actions with support from external tools, and interpreting environmental feedback and signals, all to achieve a designated goal.

As the overall operational complexity of these goals increases, a single-agent architecture may experience concrete structural bottlenecks. That is why the paradigm has progressively moved towards Multi-Agent Systems(MASs) first, in which a high-level orchestrator agent decomposes a task into smaller sub-tasks and delegates them to a sub-agent skilled enough to execute them; and then, with the emergence of open interoperability frameworks such as the Google A2A protocol and the Anthropic MCP, these MASs deployments are currently and rapidly transitioning to open-ended agentic economies. In this context, cross-boundary delegation becomes a critical operational dependency; an agent that needs to collaborate with an unknown peer must first assess the peer's competence, reliability, and security before delegating any work. However, at the moment, agents have little to no prior knowledge of unfamiliar peers and must rely on unstandardized, unverified discovery mechanisms, leading to a severe infrastructural trust deficit. Existing agentic registries, discovery protocols, and decentralized ranking algorithms fail to provide valid verifiable security, either relying on unverified self-attestation, such as Agent Cards, or by depending on unauthenticated telemetry data that is vulnerable to manipulation.

To address this infrastructure gap, this thesis presents ARBiTER, a centralized, cryptographically auditable reputation bureau designed specifically for open-ended agentic economies. The design intentionally mirrors the structural and economic macro-architecture of a traditional credit bureau, with ARBiTER serving as a neutral, ecosystem-level trust layer for agentic discovery. The system can ingest decentralized behavioral telemetry and aggregate peer interactions into a quantifiable, portable, and auctionable reputation score. To do so, ARBiTER anchors agent identity and capability claims in Self-Sovereign Identity (SSI) primitives (such as W3C's Decentralized Identity (DID), Verifiable Credentials (VC), and Verifiable Presentations (VP)); binds peer reviews to verifiable task interactions through the Delegation Agreement Protocol (DAP), a custom extension specifically for ARBiTER to the A2A protocol; and computes scores using a novel sybil-resistant aggregation algorithm, inspired by authoritative network-ranking frameworks. To ensure absolute transparency and independent auditability, ARBiTER batches incoming peer reviews into temporal units called epochs, seals them into an immutable Merkle tree data structure, and then commits them to an internal cryptographic hash-chain; the resulting trace is permanently anchored to a public, external blockchain ledger, enabling external auditors to indipendently verify the complete historical integrity of any agent's score without fully trusting ARBiTER's engine. The architectural design of ARBiTER is validated through a fully functional Proof-of-Concept (PoC) deliverable, evaluated across a series of simulated, realistic multi-agent delegation scenarios under adversarial conditions, such as a Sybil attack. The evaluation demonstrates results confirming that ARBiTER successfully isolates malicious actors, provides stable and accurate reputation convergence, and incurs minimal computational and latency overhead during the discovery phase, proving ARBiTER as a scalable, concrete architectural foundation for a trust layer in agent discovery for open-ended economies, and within the future Internet of Agents (IoA).