Collaborative Agentic AI Needs Interoperability Across Ecosystems

📝 Paper Summary

Multi-agent Agent interoperability Agentic infrastructure

The paper argues for 'Web of Agents', a minimalist architectural framework using existing web standards (HTTP, JSON) to enable interoperability between fragmented agentic AI ecosystems.

Core Problem

The agentic AI landscape is fragmenting into incompatible silos (e.g., Google's A2A, Anthropic's MCP) with distinct protocols, preventing agents from different ecosystems from collaborating.

Why it matters:

Fragmentation imposes engineering overheads, forcing developers to reimplement agents for each ecosystem or build brittle translation layers
Lack of interoperability leads to vendor lock-in, limiting user choice and stifling the innovation that comes from open, composable systems
Security risks increase when relying on complex, non-standard translation layers rather than audited, common web protocols

Concrete Example: A travel-booking agent developed by an airline using one protocol (e.g., A2A) cannot be invoked by a calendar agent running on a different ecosystem (e.g., MCP), preventing a seamless 'book flight based on my schedule' workflow.

Key Novelty

Web of Agents (Architectural Blueprint)

Proposes reusing existing, battle-tested web standards (HTTP methods, Cookies, URLs) for agent coordination instead of inventing new, heavy protocols
Introduces 'Interaction Documents'—flexible, self-describing manifests accessible via HTTP—allowing agents to dynamically discover and understand each other's capabilities without rigid schemas

Architecture

The 'Web of Agents' architectural blueprint, illustrating how agents interact using standard web technologies.

Breakthrough Assessment

7/10

A strong position paper advocating for a pragmatic shift in agent design. While it lacks empirical experiments, its proposal to align agents with HTTP/Web architecture could fundamentally solve the looming fragmentation crisis.

⚙️ Technical Details

Problem Definition

Setting: Decentralized multi-agent system coordination across heterogeneous ecosystems

Inputs: Agent capabilities, task requests, interaction context

Outputs: Standardized communication patterns and successful cross-agent task execution

Pipeline Flow

Messaging (HTTP Transport)
Interaction Specification (Interaction Documents)
State Management (Sessions/Cookies)
Discovery (Finding Agents)

System Modules

Messaging Layer

Facilitate asynchronous and structured communication using standard HTTP methods (GET/POST)

Model or implementation: Standard HTTP Protocol

Interaction Interoperability Layer

Establish shared interaction specifications so agents can understand each other's interfaces

Model or implementation: Interaction Documents

State Management Layer

Maintain conversational context and long-term history

Model or implementation: HTTP Cookies (Session) / Databases (Long-term)

Novel Architectural Elements

Rejection of 'translation layers' in favor of 'minimal standards' based on existing web infrastructure
Use of HTTP Cookies specifically for Multi-Agent short-term memory (session management)

Modeling

Base Model: Not applicable (Architectural proposal, not a specific model)

Comparison to Prior Work

vs. A2A/MCP/ACP: Web of Agents uses vanilla HTTP/Web standards rather than creating new, isolated protocol stacks
vs. Translation Layers: Proposes native interoperability via shared minimal standards rather than post-hoc bridges between incompatible systems

Limitations

The paper is a position paper and does not provide an empirical evaluation or reference implementation
Relies on the assumption that industry players will voluntarily adopt minimal standards over their own proprietary ecosystems
Does not detail the specific implementation of the 'Discovery' mechanism beyond mentioning it as a building block

Reproducibility

This is a position paper proposing an architecture. No code or trained models are associated with this work.

📊 Experiments & Results

Main Takeaways

Fragmented ecosystems (silos) threaten the scalability and utility of collaborative agentic AI by creating incompatible 'walled gardens'.
Interoperability enables security through shared scrutiny (community auditing of standard protocols) rather than obscure proprietary mechanisms.
Adopting 'minimal standards' (HTTP, Cookies, Interaction Docs) is superior to enforcing a single heavy protocol or building brittle translation layers.
Scalability is best achieved by mimicking the architecture of the World Wide Web, allowing decentralized growth without a central coordinator.

📚 Prerequisite Knowledge

Prerequisites

Understanding of RESTful web services and HTTP
Familiarity with current agent frameworks (MCP, A2A)
Basic knowledge of multi-agent systems (MAS)

Key Terms

A2A: Agent-to-Agent protocol—a specific communication standard released by Google

MCP: Model Context Protocol—an open standard by Anthropic for connecting AI assistants to systems/data

ACP: Agent Communication Protocol—a protocol released by IBM for agent interaction

HTTP: Hypertext Transfer Protocol—the foundation of data communication for the World Wide Web, proposed here as the transport layer for agents

Interaction Document: A proposed flexible document format (similar to an API spec) that agents use to describe their inputs, outputs, and capabilities to others

JSON-RPC: A remote procedure call protocol encoded in JSON; the paper argues against enforcing this as a mandatory standard in favor of simple HTTP

OAuth: An open standard for access delegation, commonly used for token-based authentication on the web