TL;DR
This paper introduces Agent-OSI, a new architecture for a decentralized "Internet of Agents" (IoA) that enables secure, interoperable communication and pay-per-use settlement. By leveraging existing web standards like HTTP 402, Agent-OSI aims to break down platform silos and pave the way for autonomous agents to interact seamlessly across different systems.
What Problem Does This Paper Address?
The rapid advancement of Large Language Models (LLMs) is driving the evolution from an "Internet of Information" to an "Internet of Agents" (IoA). In this future, autonomous agents will discover services, negotiate terms, execute tasks, and exchange value with minimal human intervention. However, current agent deployments are often confined to proprietary platforms, lacking a common framework for interoperability, trust, and efficient pay-per-use settlement. This fragmentation hinders the potential of a truly decentralized IoA, where agents can reliably interact across administrative domains. Agent-OSI aims to bridge this gap by providing a standardized, layered architecture.
Methods
Agent-OSI is a functional interoperability architecture, not a new network protocol. It's inspired by the layered approach of the OSI model but adapted for the IoA. The core contributions are in agent-to-agent (A2A) communication (Layer 2) and a backend-agnostic settlement protocol (Layer 4) built on HTTP 402 (Payment Required).
- Layered Architecture: Agent-OSI defines six layers:
- L1: Connectivity & Secure Transport: Leverages existing Internet protocols like IP, TCP, and TLS/QUIC.
- L2: A2A Messaging & Routing: Provides asynchronous messaging with inbox semantics, routing, and offline delivery capabilities.
- L3: Identity, AuthN & Authorization: Aligns with decentralized identifiers (DIDs) and Web authorization standards like OAuth 2.0.
- L4: Settlement & Metering: This is a key innovation, using HTTP 402 as a challenge-response primitive for pay-per-use services. It supports quotes, receipts, and escrow-based settlement.
- L5: Verifiable Execution & Provenance: Enables agents to provide portable evidence of execution and provenance, crucial for trust and dispute resolution.
- L6: Semantic Interop & Orchestration: Handles capability discovery, tool binding, and workflow composition.
- HTTP 402 for Settlement: The paper repurposes HTTP 402, traditionally "Payment Required," as a standardized way for agents to request payment for services. This allows clients to pay per request without proprietary billing APIs, while supporting replay protection and auditability.
- Backend Agnosticism: The settlement layer is designed to be pluggable, supporting various backends like escrow contracts, payment channels, or signed off-chain receipts. The prototype uses a blockchain escrow.
- Cross-Layer Bindings: Explicit, signed artifacts (quotes, receipts, provenance) are used to bind different layers, preventing equivocation between what was executed and what was paid for.
Datasets
Not discussed in the paper. The evaluation uses a self-implemented prototype and synthetic workloads.
Key Insights and Ideas
- HTTP 402 as a Universal Payment Primitive: Repurposing an existing HTTP status code for agent settlement is a clever way to leverage web infrastructure and avoid introducing entirely new protocols. It provides a standardized, application-layer mechanism for pay-per-use interactions.
- Decoupling Negotiation/Delivery from Settlement: Keeping negotiation and delivery logic off the core settlement backend significantly reduces per-session costs. This is a crucial insight for making decentralized agent interactions economically viable.
- Verifiable Provenance as the Trust Anchor: Instead of relying on a central platform, trust is built through cryptographically verifiable evidence of execution and provenance, which can be used for settlement and auditing.
- Interoperability Through Standardized Artifacts: The architecture emphasizes standardized, signed artifacts (quotes, receipts, provenance) as controlled cross-layer interfaces. This allows for substitutability of underlying mechanisms without breaking the overall system.
Key Results
The paper presents a prototype called AgentMarket and evaluates its performance.
- Cost Reduction: Agent-OSI reduces per-session on-chain settlement costs by approximately 51% compared to a more on-chain baseline in their prototype.
- Latency Dominated by Task Execution: For generative workloads (like image generation), the end-to-end latency is primarily determined by the task execution time, not the settlement confirmation.
- Throughput: The system achieves a steady-state message throughput of around 50 messages/second and settlement throughput of about 40 transactions/second, with completed sessions peaking at around 20 sessions/second. This throughput is currently limited by the underlying blockchain settlement rate.
Limitations and Future Work
- Privacy and Linkability: A tension exists between the cross-layer bindings needed for replay protection and fair exchange, and the desire for privacy and linkability. Future work includes developing privacy-preserving messaging overlays, selective disclosure receipts, and provenance minimization techniques.
- Identity and Key Management: Scaling decentralized identity, delegation, and key management for agents remains an open challenge.
- Settlement Semantics and Dispute Resolution: Precise definitions for metering, robust dispute interfaces, and handling of various settlement backends are areas for further development.
- "Verifiable Execution" Definition: Standardizing evidence statements, schemas, and scalable verification strategies for different assurance levels of verifiable execution is needed.
- Wide-Area Deployment: The current prototype evaluation is on a local testbed. Future work includes evaluating Agent-OSI in a wide-area, public-chain, and heterogeneous-trust-domain environment.
- Closing Architectural Gaps: Features like client-side encryption for access-gating delivery are architectural options not enforced by the current prototype.
Summary updated