New Protocols Aim to Empower AI Agents to Seamlessly Navigate Digital Tasks

AI Agents Struggle to Manage Complex Digital Environments

Many companies are launching AI agents that can perform tasks on users' behalf, such as sending emails, editing documents, or managing databases. Despite their promise, these agents have faced challenges in effectively interacting with the diverse components of our digital lives. The lack of mature infrastructure to support these interactions is a significant hurdle.

The Role of Protocols Like MCP and A2A

Anthropic and Google have introduced protocols to tackle these challenges. Anthropic developed the Model Context Protocol (MCP), which standardizes how AI agents communicate with various programs by translating between natural language and code. MCP has gained popularity quickly, with over 15,000 servers supporting it.

Google introduced the Agent2Agent protocol (A2A), which focuses on governing the interactions between AI agents themselves. A2A aims to enable more sophisticated, multi-agent collaborations by establishing clear rules on what agents must do, should do, and must not do, ensuring safe and effective communication. Many companies, including Adobe and Salesforce, have partnered with Google to adopt A2A.

MCP and A2A complement each other, with MCP facilitating information exchange and A2A managing agent-to-agent communication.

Security Concerns and Challenges

Security remains a critical concern. AI agents can be vulnerable to attacks, such as indirect prompt injection, which can hijack an agent to perform malicious actions like leaking private data. Currently, MCP lacks built-in security design, raising concerns about the potential risks of empowering AI agents with greater control.

Experts like Bruce Schneier warn about the security risks as AI agents gain more real-world power. However, researchers like Zhaorun Chen see potential for protocols like MCP and A2A to evolve with better security features, possibly similar to internet protocols like HTTPS. Standardization may also aid in identifying and mitigating vulnerabilities more effectively.

Openness and Governance of Protocols

Multiple organizations are developing protocols for AI agents, including Cisco, IBM, and academic groups. Open-source development is common, as seen with MCP and A2A, facilitating faster and more transparent progress. Google donated A2A to the Linux Foundation to ensure collaborative governance, while MCP remains owned by Anthropic but is licensed permissively.

There is ongoing discussion about governance models to ensure these protocols serve broad interests, with possibilities for more open, community-driven decision-making in the future.

Efficiency and Natural Language Interfaces

Both MCP and A2A use natural language to communicate between agents and services, leveraging the AI models' native way of processing information. This approach simplifies integration but comes with trade-offs.

Natural language communication is less precise than traditional APIs and requires more computational tokens, increasing costs and inefficiencies. For example, an AI agent summarizing a document via another service must send and receive full text multiple times, doubling token usage.