Multi-Agent Wiki
Protocol

Protocol-mediated Agent Network

Connect tools, agents, clients, and platforms via standard protocols — MCP, A2A, ACP, Agent Client Protocol.

Live visualizationAnimated topology — press Space to play / pause
·Diagram
Live
1 / 8

—

Definition

Use standard protocols — MCP, A2A, ACP, Agent Client Protocol — to connect tools, agents, clients, and platforms across frameworks and vendors.

Category: Protocol interconnect

When to use

Cross-framework interop, enterprise integration, IDE-to-coding-agent, tool ecosystem standardization.

When not to use

Monolithic demos, environments where every tool is local and no standardization is required.

How to implement

  1. MCP is for agent-to-tool/resource calls — don't reuse it for agent-to-agent communication.
  2. A2A / ACP handles agent discovery, tasks, messages, and collaboration.
  3. The Agent Client Protocol covers IDE/client to coding-agent links.
  4. All protocol boundaries require auth, authorization, audit, and rate limits.

Minimal pseudocode

TypeScript
interface AgentRuntimePorts {
  mcp: MCPClient;                 // tools / resources / prompts
  a2a: AgentDirectoryClient;      // remote agents
  client: AgentClientServer;      // IDE / web / CLI
  events: EventBus;               // observability
}

Recommended trace events

  • protocol.mcp.tool_call
  • protocol.a2a.task.created
  • protocol.client.session.started
  • protocol.auth.failed

Common failure modes

  • Conflating MCP, A2A, and Agent Client Protocol into a single protocol.
  • External agents granted too much permission.
  • No identity or audit.

Implementation checklist

  • Input/output schemas defined.
  • Each agent's permission boundary defined.
  • Every agent call carries a run id / trace id.
  • Failure, timeout, cancel, and retry strategies defined.
  • Context passed is the minimum required, not the full history.
  • High-risk actions are gated by approval or a verifier.

References