Why Are Agent Protocols Like MCP and A2A Needed?

The topic of AI agents is rapidly evolving. You’ve probably heard about Anthropic’s Model Context Protocol (MCP), which grew aggressively from 500 MCP servers in February to over 4,000 as of today. As the community was still figuring out MCP, Google announced its contribution to the agent communication spaces by releasing Agent-2-Agent (A2A). The announcement was met with excitement and confusion.

This rapid proliferation raises questions: Why do we need multiple agent protocols? Are MCP and A2A overlapping solutions, or do they complement each other? And critically, what tangible ROI can organizations expect from adopting these emerging AI agent technologies? This article will explore these questions in depth.

Agent Communication Basics

Let’s start with the fundamentals. While definitions vary, a practical way to describe an AI agent is “software that uses AI models to perform tasks on behalf of users.”

Typically, agents mediate interactions between two main entities: consumers (humans or other agents requesting information or actions) and data sources or external systems (providing information or executing actions). The agent leverages generative AI models as an intelligent intermediary, translating user requests into executable tasks and coordinating responses or actions from external systems.

Anthropic’s MCP standardizes how AI models communicate with external tools, databases and APIs. It provides models with clear context about what external services can do and what data they need. With MCP, models can independently select and execute tasks more efficiently. Think of MCP as a standardized connector, like USB for data transfer. Just as USB made connecting devices simpler and more reliable, MCP makes interactions between models and external systems smoother and easier to manage.

Google’s A2A works at a higher level, enabling smooth interactions between multiple agents. It ensures secure communication, manages the state and context of interactions, helps agents negotiate tasks with each other, and makes it easy for agents to discover one another. Essentially, A2A orchestrates collaboration across multiple specialized agents, coordinating their actions to achieve complex goals efficiently.

Let’s use a simple example demonstrating how A2A and MCP work together. Suppose you ask an agentic multi-agent system to work on the following prompt: “Book me a trip to Paris with warm weather, outdoor activities and a maximum budget of $3,000.

Here’s what happens next:

• A2A coordinates specialized sub-agents to handle individual tasks:
  • Flights: Google Flights agent
  • Weather: Weather Channel agent
  • Hotels: Booking.com agent
  • Activities: TripAdvisor agent
• Each sub-agent uses MCP to interact efficiently with its specific external data sources.
• A2A then gathers the results from these sub-agents, compares available options, selects the best choices based on your requirements and assembles a complete travel itinerary for you.

In short, MCP ensures that each agent smoothly retrieves the necessary data, while A2A manages the overall collaboration and delivers an optimal, unified solution to you.

Do We Need These Protocols Anyway?

Protocols should be used because we need them, not just for the sake of it. In a recent panel that I moderated during the Rootly AI Labs launch event, Pete Koomen, general partner at Y Combinator, noted that while they’ve built many AI agents to help run Y Combinator, none of them rely on MCP or A2A. Do we need these protocols?

Yoko Li, a partner at Andreessen Horowitz, believes AI models have a bright future and will eventually reshape our industry by doing a lot on our behalf. But as of today, the technology isn’t fully mature yet. Models forget information, get confused with the tools at their disposal and how to use them, which is why we need these procedure layers to help them get there. Protocols like MCP and A2A currently play this crucial role as “hand-holding” tools, assisting agents and the models behind them to navigate tasks they cannot handle independently.

And while simple agentic workflows might not need them, sophisticated ones, especially those involving multiple AI agents. Miku Jha, applied director at Google Cloud and co-creator of A2A, emphasized that enterprises will adopt agent-driven workflows only if they trust agent technologies to be reliable, similar to existing enterprise tools. While generative AI is impressive in many aspects, reliability isn’t one of its strong attributes. That was the intent behind A2A. Google’s team collaborated closely with enterprise partners to understand their technical challenges around building sophisticated agent workflows, and created the protocol to address these real-world issues.

Measuring ROI in the Agent Era

Companies typically leverage agents in two primary ways: modernizing existing workflows or enabling new user interactions with products. Once identified, companies must evaluate whether agent-generated outputs meet their quality standards and translate into tangible cost savings or increased revenue. Jha emphasized that companies often mistakenly start their exploration by selecting specific tools, like MCP or A2A, rather than clearly defining their use cases and criteria for success. She advised reversing this approach: Companies should first identify their use cases and what success means in terms of business value before investing heavily in agent technologies.

However, monitoring and measurement for agent-driven systems remain underdeveloped. Even basic reliability concepts, such as measuring “how many nines” of uptime AI systems deliver or defining SLAs, are still nascent, and evaluating GenAI output is tricky. During the last SRECon Americas, the audience was shocked to learn that Microsoft was using NPS (Net Promoter Scores) to measure the performance of AI models.

Conversely, Li highlighted scenarios where traditional ROI measurement might fail or simply not make sense. She cited AI companions as a vast market and as an example of consumers who might be in a situation where they didn’t know they needed an AI companion, but now could not live without it — how do you measure ROI in that case?

Building Is the Way Forward

AI agent technology is still in its early days, yet companies are enthusiastically exploring its potential and eager to seize first-mover advantages. Koomen highlighted that we stand at “one of the most interesting moments in human history,” at the beginning of a technological revolution.

While these protocols may not last, MCP adoption has been rapid, and even Anthropic’s competition, such as OpenAI, has adopted it. Meanwhile, Microsoft recently announced that it is adopting Google’s A2A for its inter-agent communication needs. These protocols will help bring the agent to the next level of reliability, which may enable companies to turn prototypes into projects that bring value. With only about 5% of GenAI projects turning into profitable products, there is still a long way to go.