Argmin AIArgmin AI
About usOptimize your agent
RLHFBasic RAGPrompt OptimizationAgentic RAGQuantizationSafety BenchmarksLLM-as-a-JudgeChain-of-ThoughtMulti-Agent SystemsFunction CallingEfficient InferenceHallucination Detection

Practical survey and roadmap for four agent interoperability protocols (MCP, ACP, A2A, ANP)

Agentic AIAgent CommunicationMulti-Agent SystemsAgent OrchestrationSurvey
May 4, 20256 min

Overview

Decision SnapshotNeeds Validation

The paper synthesizes protocol specs and community docs; recommendations are practical but based on specification analysis rather than large-scale field studies.

Citations3

Evidence Strength0.70

Confidence0.78

Risk Signals9

Trust Signals

Findings with numeric evidence: 4/4

Findings with evidence refs: 4/4

Results with explicit delta: 0/0

Reproducibility

Status: No open assets linked

Open source: Partial

At A Glance

Cost impact: 70%

Production readiness: 70%

Novelty: 40%

Authors

Abul Ehtesham, Aditi Singh, Gaurav Kumar Gupta, Saket Kumar

Links

Abstract / PDF

Why It Matters For Business

Standardizing agent interfaces reduces engineering cost, improves security, and enables reusable agent services across teams and vendors.

Who Should Care

CTOProduct ManagerML EngineerEngineering LeadData Scientist

Summary TLDR

This survey compares four emerging protocols for LLM-powered agents: MCP (JSON-RPC for secure tool/context injection), ACP (RESTful, MIME-typed multimodal messaging and session handling), A2A (agent-to-agent capability exchange using Agent Cards and SSE/push updates), and ANP (decentralized DID-based peer discovery and JSON-LD descriptions). It catalogs architectures, transport options, discovery methods, and security threats, and proposes a four-stage adoption roadmap: start with MCP for tool access, add ACP for structured messaging, adopt A2A for enterprise agent workflows, and extend to ANP for open, decentralized agent markets.

Problem Statement

LLM-driven agents need standardized ways to discover capabilities, share structured context, and securely coordinate actions. Current integrations are ad-hoc, hard to scale, and vary across vendors, creating security gaps, brittle adapters, and costly engineering work.

Main Contribution

Survey and concise comparison of four protocols: MCP, ACP, A2A, ANP

Detailed protocol descriptions (architecture, transports, discovery, security) and lifecycle threat tables

Key Findings

Four distinct protocols target different interoperability layers

Numbers4 protocols compared (MCP, ACP, A2A, ANP)

Practical UsePick the protocol that matches your scope: MCP for model↔tool, ACP for typed agent messaging, A2A for enterprise task delegation, ANP to go decentralized.

Evidence RefAbstract; Table 7

A practical adoption sequence reduces integration risk

Numbers4-stage roadmap (MCP → ACP → A2A → ANP)

Practical UseDeploy incrementally: enable tool invocation first, add structured messaging, then cross-agent workflows, and finally open-network discovery.

Evidence RefSection 9 (Phased Adoption Roadmap)

What To Try In 7 Days

Prototype MCP for model-to-tool calls and centralize tool manifests

Sign and version one Agent Card/manifest and host under /.well-known

Add TLS + JWS-signed message parts to an existing HTTP-based agent endpoint

Agent Features

Memory
Session context / session IDsManifest and Agent Card metadata
Planning
Task delegationSession-aware workflowsCapability negotiation
Tool Use
Tool invocation (model-controlled tools)Typed function calls (JSON schemas)
Frameworks
JSON-RPCHTTP/RESTJSON-LDW3C DIDsMIME multipart messagesServer-Sent Events (SSE)
Is Agentic

Yes

Architectures
client-server (MCP)brokered HTTP/REST (ACP)peer-like client↔remote (A2A)decentralized P2P with DIDs (ANP)
Collaboration
Agent discoveryCapability / Agent Card exchangeArtifact exchange

Optimization Features

System Optimization
Stateless default for ACP to enable scalable load balancing

Reproducibility

Code AvailableNo
Data AvailableNo
Open Source StatusPartial
LicenseUnknown

Risks & Boundaries

Limitations

Survey synthesizes protocol specs and early docs; lacks large-scale empirical deployments

Protocol specs and adoption status are evolving rapidly; details may change

When Not To Use

If you run a single monolithic app with no external tool calls, protocol adoption may add unnecessary complexity

Do not adopt decentralized ANP if you require strict centralized control and auditability

Failure Modes

Manifest or Agent Card spoofing leading to impersonation

Tool or prompt injection causing unsafe model behavior

Context Entities

Models

Function calling (OpenAI)RAG (retrieval-augmented generation)

You May Also Want to Read

Argues that 'agentic' buzzwords mostly rebrand decades-old agent and multi-agent research

0.60
0.30
0.50
0

Calling LLM-based systems 'agentic' can hide reusable MAS solutions; adopting MAS standards and algorithms reduces engineering risk and speeds integration.

Key finding

'Agentic AI' and 'Multiagentic' are mostly new labels for existing concepts (intelligent agents and multi-agent systems).

Numbers: Dozens of LLM-agent frameworks emerged 2023–2025

TRiSM: practical trust, risk and security controls for LLM-based multi-agent systems

0.60
0.40
0.70
0

Agentic systems increase autonomy and regulatory exposure; TRiSM reduces legal, reputational and operational risk while enabling auditable, compliant deployment.

Key finding

Academic interest in agentic AI has exploded, especially after ChatGPT's launch.

Numbers: Multi-agent papers: 890 (2019) → 18,500 (2024); LLM-agent papers: ~0 → 9,800 (post-2022)

RAPS: intent-driven, reputation-aware publish–subscribe for adaptive multi-agent LLM coordination

0.60
0.70
0.60
0

RAPS offers inference-time, training-free coordination that improves task accuracy, scales with more agents, and reduces single-point failures — useful for teams building modular LLM services and open agent marketplaces.

Key finding

RAPS achieves strong end-task gains across five benchmarks.

Numbers: 90.0% average accuracy/Pass@1 across MMLU, GSM8K, SVAMP, AQuA, HumanEval (Table 1).

ACP: a layered, federated protocol for secure cross-platform agent-to-agent collaboration

0.60
0.70
0.60
0

ACP lets organizations stitch together agents from different vendors into secure, auditable workflows. That reduces manual coordination and enables cross-company automation while preserving regulatory controls.

Key finding

ACP reduces inter-agent communication latency compared to plain JSON-RPC over HTTPS (author claim).

Numbers: 40% reduction (claimed in abstract); Table 3 shows ACP avg latency 58 ms vs JSON-RPC 145 ms

A benchmark showing LLMs can coordinate by reading environments but struggle at partners' beliefs and joint planning

0.30
0.60
0.40
10

LLMs can act as zero-shot coordination partners for tasks where the environment dictates the correct action (logistics routing, scripted multi-robot tasks), cutting training time; but they are unreliable when partner modeling or multi-step joint planning is required.

Key finding

LLM agents match or exceed RL on environment-driven Overcooked layouts.

Numbers: GPT-4-turbo: 260 (AA layout) vs PBT: 190 (Table 1)