IETF agentproto Begins — The IDL Standard for AI Agents to Negotiate Directly

We may soon see a future where AI agents created by different companies communicate directly and negotiate tasks among themselves. The Internet Engineering Task Force (IETF), which sets the fundamental rules for the internet, has recently begun full-scale operations for the 'agentproto' project to create communication standards for AI agents. We’ll break down these new standards—which aim to let agents from all over the world communicate freely across a single, vast network, regardless of platform barriers—in an easy and engaging way.

What is the Interface Description Language (IDL), the AI Agent's Agreement?

The most noteworthy standard proposal from this meeting is the 'Interface Description Language (IDL)' information model. IDL is an independent rule model that is not tied to any specific platform, serving as a way for AI agents to explain to other agents what they can do and what constraints they have.

It’s similar to how two agencies might negotiate specific terms before signing a business contract. Without human intervention, AI agents use this standardized agreement to verify each other's capabilities and operational limits. For example, they can clearly exchange constraints like, "My available budget is a maximum of $50, and human final approval is required before payment."

Once these common protocols are in place, AIs from different companies can start working together safely the moment they meet. Developers won't have to struggle to build custom adapters every time they need agents to communicate. Just as Internet Explorer and Chrome operate based on the same web standards, a path is opening for AIs to cooperate freely on a unified specification.

A Standard Specification to Unify Disparate LLM Streaming

When building AI services, a streaming feature where answers appear on the screen in real-time, character by character, is now essential. However, developers have faced significant hurdles in handling this smoothly. This is because every framework or AI model provider—such as LangChain, Vercel AI SDK, or LiteLLM—has had different API specifications for sending and receiving streaming data.

The standard proposal introduced at this IETF meeting aims to resolve this fragmentation. The core idea is to unify the LLM streaming communication format based on Server-Sent Events (SSE) technology. To this end, a new standardized data format has been proposed called application/llm-request+json.

If this standard is widely adopted, developers will no longer be dependent on the proprietary API specifications of specific frameworks or models. Because they can build consistent connection infrastructure just as they follow web standards, the inefficiency of having to rewrite streaming reception code every time they switch AI models or frameworks will be significantly reduced.

Empowering Agent Collaboration with Real-Time Data Transfer

The Model Context Protocol (MCP), led by Anthropic and Microsoft and recently updated to version 2.0, is also gaining momentum from this IETF standardization trend. At this meeting, specific proposals were discussed to run MCP on a faster and more powerful real-time communication specification—specifically, by incorporating Media over QUIC Transport (MoQT), a technology optimized for real-time data streaming.

MoQT is a high-speed transmission technology originally designed for use cases like video streaming that require extremely low latency. When combined with agent communication standards, AI agents will be able to go beyond exchanging static text messages and synchronize large amounts of data or complex, real-time context without delay.

It’s like moving from walkie-talkies where you have to wait for a turn to speak, to being connected by high-speed dedicated lines. Once this standard takes hold, a robust, internet-grade backbone network that allows agents worldwide to collaborate in real-time across company and platform boundaries will become a reality.

Breaking Down Corporate Barriers to an Open Agent World

Just as web browsers began to follow standard specifications and opened up the vast web world we have today when the internet first appeared, AI agents have begun to walk the same path.

Agents that were previously locked within the closed platforms of specific companies are now preparing to connect under common internet rules. The foundation of a true 'Agent Internet' is being laid—not a proprietary ecosystem dominated by a single company, but one where anyone can freely participate and interoperate.

Let’s keep a close and exciting watch on what practical changes these new standards will bring toward a future where diverse agents converse and cooperate as naturally as humans!