The Protocol Layer

How Agents Will Talk to the World

In early 2024, a large enterprise software company -- one you've used -- had 23 engineers working on what they called the "connector layer." Their AI assistant needed to talk to Salesforce, Jira, Slack, Google Calendar, SAP, ServiceNow, Confluence, GitHub, and about forty other tools their enterprise customers depended on. Each connector was custom. Each one required a dedicated engineer who understood that tool's API, its authentication model, its rate limits, its pagination scheme, and its data model.

The math was ugly. Forty-eight connectors. Average maintenance cost of 15 engineering hours per connector per month. That's 720 engineering hours a month -- roughly 4.5 full-time engineers doing nothing but keeping existing integrations alive. Not building new features. Not improving the product. Maintaining plumbing.

Then, in November 2024, Anthropic released the Model Context Protocol specification. By March 2025, OpenAI announced they were adopting it for ChatGPT. By mid-2025, MCP had thousands of community-built servers. When your competitor adopts YOUR protocol, something structural has happened.

MCP, A2A, and the Coming Protocol Standardization

In the early 1990s, connecting two computers required knowing both systems' proprietary networking protocol. Then TCP/IP won. Not because it was technically superior to every alternative. Because it was open, and openness created a network effect that no proprietary protocol could match. HTTP sat on top and did the same thing for the web. MCP and A2A are doing the same thing for agents.

Think of it like electrical outlets. Before standardization, every appliance manufacturer designed their own plug shape. Standardized sockets meant any appliance from any manufacturer worked in any outlet. MCP is the standard outlet for AI tools. A2A is the standard plug for AI agents.

MCP uses a client-server architecture built on JSON-RPC 2.0. The model runs an MCP client. Each tool runs an MCP server exposing three capabilities: Tools (actions the agent can take), Resources (data the agent can read), and Prompts (pre-built instructions the server offers).

A2A solves the agent-to-agent problem. When Agent A needs to delegate to Agent B -- built by a different company, running on a different model -- A2A standardizes the handoff. Every A2A-compatible agent publishes an Agent Card at /.well-known/agent.json describing what it can do and what authentication it requires.

The company list at A2A's launch tells you where this is heading: Salesforce, SAP, ServiceNow, Atlassian, Box, Cohere, and roughly 50 others. These aren't AI startups chasing a trend. They're enterprise software incumbents who see agent interoperability as structural.

Why Your Custom Integrations Have a Two-Year Shelf Life

The integration cost cliff. An enterprise AI product connecting to 15 tools carries 15 custom integrations, each requiring dedicated maintenance. With MCP, the product carries one MCP client and points it at 15 community-maintained servers. When Salesforce updates their API, the Salesforce MCP server maintainer updates the server. Your product doesn't change.

Vendor lock-in erosion. If your competitive advantage is "we integrate with 200 tools," MCP eliminates that moat. When any product can access any MCP server, integration breadth stops being a differentiator. The new differentiator becomes what you DO with the tools -- orchestration quality, context management, reliability.

Multi-vendor agent orchestration. A2A makes it possible to compose agent systems from multiple vendors without building custom middleware. A procurement workflow might involve your internal agent on Claude, your ERP vendor's agent on Gemini, and your logistics partner's agent on a fine-tuned open-source model.

Migrating to a Protocol-Based Architecture

Consider a mid-size B2B SaaS company -- call them Meridian -- that sells a workflow automation platform. 340 enterprise customers, 52 tool integrations, 12-person integration engineering team. Monthly maintenance: 800 engineering hours.

The 5x reduction in maintenance hours is real but not the primary value. The primary value is the reallocation: 9 engineers building product instead of 4. And the A2A gateway -- Meridian's Agent Card -- creates a new acquisition channel that didn't exist before the protocol layer.

Remember This

1. MCP standardizes agent-to-tool connections (vertical) and A2A standardizes agent-to-agent communication (horizontal). They're complementary, not competing. Together they form the protocol layer.

2. OpenAI adopting Anthropic's protocol, and Anthropic donating it to the Linux Foundation, are structural signals that MCP is becoming infrastructure, not a product. Products built on proprietary connectors need to plan their migration now.

3. Protocols shift the competitive surface. When connectivity is solved, the differentiator becomes what you DO with the connections: context quality, harness design, evaluation rigor, governance enforcement.