Survey: Agent Communication Protocols in 2026

Abstract

The AI agent ecosystem has produced multiple competing communication protocols in the span of 18 months. This survey analyzes five prominent protocols — MCP, A2A, ACP, AG-UI, and ANP — comparing their design goals, transport mechanisms, capability models, and trust assumptions. We conclude that these protocols are largely complementary rather than competing, and identify the contract layer gap that OSSA fills above them.

1. Introduction

Between 2024 and 2026, five significant agent communication protocols emerged:

1. MCP (Model Context Protocol) — Anthropic, Nov 2024
2. A2A (Agent-to-Agent) — Google, Apr 2025
3. ACP (Agent Communication Protocol) — IBM/BeeAI, 2025
4. AG-UI (Agent-User Interaction Protocol) — CopilotKit, 2025
5. ANP (Agent Network Protocol) — ANP Working Group, 2025

Each addresses a different aspect of agent communication, but none provides a complete solution. Understanding their boundaries is essential for building interoperable agent systems.

2. Protocol Comparison

2.1 Design Goals

2.2 Transport Layer

2.3 Capability Model

MCP defines three capability types:

• Tools: Functions the LLM can invoke
• Resources: Data the LLM can read
• Prompts: Template interactions

A2A defines capabilities through Agent Cards:

• Skills: What the agent can do
• Input/Output schemas: Typed message formats
• Streaming support: Long-running task updates

ACP focuses on collaboration:

• Message types: Request, response, notification
• State management: Shared context across agents
• Group communication: Multi-party conversations

AG-UI focuses on UI integration:

• Event streams: Text, tool calls, state updates
• Lifecycle events: Start, progress, complete, error

ANP focuses on identity:

• DID-based addressing: Every agent has a decentralized identifier
• Capability attestation: Cryptographically signed capability claims

3. Analysis

3.1 The Layer Diagram

These protocols operate at different layers of the agent communication stack:

 Application Layer    [Agent Logic / Orchestration]
         |
 Contract Layer       [OSSA — portable agent definitions]    <-- GAP
         |
 Communication Layer  [MCP | A2A | ACP | AG-UI | ANP]
         |
 Transport Layer      [HTTP | WebSocket | stdio | P2P]
         |
 Network Layer        [TCP/IP | DNS]

No protocol addresses the Contract Layer: the portable, declarative definition of what an agent is, what it can do, and what governance rules it operates under.

3.2 Complementarity vs. Competition

A common misconception is that these protocols compete. In practice:

• MCP + A2A: MCP handles tool invocation; A2A handles agent-to-agent delegation. An agent uses MCP to call tools and A2A to delegate to other agents.
• A2A + ACP: A2A handles bilateral handoffs; ACP handles multi-party collaboration with shared state.
• AG-UI + MCP: AG-UI streams agent activity to a frontend; MCP connects the agent to backend tools.
• ANP + DUADP: ANP provides DID-based identity; DUADP provides DNS-based discovery. They can coexist.

3.3 What's Missing

All five protocols assume the agent is already defined and running. None answers:

• How do I define an agent portably? (Identity, role, capabilities, governance)
• How do I move an agent between platforms? (Vendor-neutral contract)
• How do I validate an agent before deployment? (Schema validation, policy checking)
• How do I map agent behavior to compliance frameworks? (SP 800-53, SOC 2)

This is the contract layer that OSSA provides.

4. OSSA's Position

OSSA operates above the communication protocols and below the application logic:

An OSSA manifest can declare support for multiple protocols:

apiVersion: ossa/v0.4.0
kind: Agent
metadata:
  name: code-reviewer
spec:
  capabilities:
    - protocol: mcp
      tools: ["read_file", "write_file"]
    - protocol: a2a
      skills: ["code-review", "security-audit"]
  discovery:
    duadp:
      enabled: true

5. Adoption Data

Based on our survey of 200 engineering leaders (January 2026):

MCP has the strongest adoption, driven by Anthropic's Claude Code integration and broad IDE support. A2A is growing rapidly thanks to Google's ADK integration.

6. Recommendations

1. Use MCP for tool integration — it's the most mature and widely supported
2. Use A2A for agent delegation — Google's backing ensures longevity
3. Use OSSA for agent definition — the contract layer is protocol-agnostic
4. Use DUADP for discovery — DNS-native, works with existing infrastructure
5. Don't pick one protocol exclusively — the stack is designed to be composable

7. Conclusion

The agent communication protocol landscape is converging, not fragmenting. Each protocol addresses a distinct concern, and they compose well together. The remaining gap is the contract layer: a portable, declarative way to define what an agent is, independent of how it communicates. OSSA fills this gap.

---

References:

• Anthropic. "Model Context Protocol Specification." modelcontextprotocol.io, 2024.
• Google. "Agent-to-Agent Protocol." github.com/google/A2A, 2025.
• IBM Research. "Agent Communication Protocol." github.com/i-am-bee/beeai, 2025.
• CopilotKit. "AG-UI Protocol." docs.ag-ui.com, 2025.
• ANP Working Group. "Agent Network Protocol." agentnetworkprotocol.com, 2025.
• NIST. "Request for Information: CAISI." Docket NIST-2025-0035, 2026.