Retirement Gift for the Father of TCP/IP: Issuing ID Cards to AI Agents

At 83 years old, Vint Cerf retired from Google just last week. Less than a week later, TechCrunch revealed his next step: leading the creation of an identity standard for AI agents on the open Internet. The supporting implementation plan, already presented by Innovation Labs, is called DNSid—linking every agent to an existing Internet domain and recording its registration trail via cryptographic proofs.

At first glance, that sounds like a post‑retirement side project. But if you’ve built anything “agent‑related” in 2026, you’ll immediately see what he’s aiming at.

What problem is he actually solving

Cerf’s own words:

“This was largely triggered by the rise of AI agents — what rights do they have, where do those rights come from, who is responsible for what they do in a given context, how and where is identity established, and on what basis do you trust it?”

This isn’t an academic question. The moment you put an agent on the public Internet to book flights, handle approvals, or call APIs on your behalf, problems pile up instantly:

• How does a target service know the request comes from an agent you authorized, not an impostor?
• If an agent claims today to have permission to call a tool but loses that permission tomorrow, how does the server perceive the change?
• When something goes wrong, how can we trace responsibility—model vendor, deployer, or end user?

Right now, each platform issues its own tokens, signs its own keys, and maintains its own whitelist. It works at small scale but collapses once it grows. DNSid aims to fix this: anchoring identity to the Internet’s existing trust infrastructure (DNS + public‑key cryptography), making identity verifiable, traceable, and revocable—without a central arbitrator.

Anyone familiar with TCP/IP will recognize the logic: use the smallest common envelope, avoid central control, and let implementers compete below the layer.

Cerf is blunt about “natural language as protocol”

Another quote from Cerf at the Open Frontier Conference deserves to be pinned to every agent team’s wall:

“I don’t think English is the best choice. English is flexible but also ambiguous. Precision between interacting agents is very important—one agent must ensure the other completely understands the agreement they just reached.”

He compared it to the “telephone game”: after ten passes, the original sentence is unrecognizable. Now imagine an automated chain of hundreds of agents, each hop interpreting natural language anew—the error amplifies at every layer, producing not collaboration but collective hallucination.

This runs counter to a strong industry undercurrent. The default narrative over the past two years has been “natural language as universal glue,” because LLMs are powerful, low‑friction to develop, and look great in demos. Yet once real multi‑agent orchestration or cross‑vendor cooperation begins, engineers quietly push key interactions back into structured schemas, JSON‑RPC, gRPC, or even fully formal contracts. Cerf just said the quiet part out loud.

The race is already on: A2A and MCP are in the lane

Seen in context, Cerf’s rallying call doesn’t start from zero; the standards battle has been underway for more than a year:

• Google Cloud’s A2A (Agent‑to‑Agent) protocol — released April 2025, 50+ partners, hosted by the Linux Foundation. It targets agent discovery, negotiation, and cooperation.
• Anthropic’s MCP (Model Context Protocol) — released November 2024, over 97 million downloads. It standardizes channels between models and tools/data sources rather than pure agent‑to‑agent links, yet addresses the same core issue: standardized capability declaration and invocation.
• DNSid / Cerf route — a deeper layer, concerned solely with identity. Think of A2A as HTTP, MCP as USB, and DNSid as DNS + TLS certificates.

Most likely these three won’t replace each other but form a layered stack—identity at the bottom, communication in the middle, capability negotiation on top. Whoever gains influence across all three layers will hold the Cisco + VeriSign + IETF‑level position in the Agent Age.

China is already moving

Few noticed that on May 8 the Cyberspace Administration of China, NDRC, and MIIT jointly issued the Implementation Guidelines for Normative Application and Innovative Development of Intelligent Agents. One passage almost pre‑empts Cerf’s vision in policy form:

Explore establishing an agent registration platform to provide services such as digital‑identity management, discovery, and capability declaration... Research foundational technologies for agent identification, trusted interconnection, compliant payment, security protection, and conflict resolution.

It even names a specific protocol—AIP (Agent Interconnection Protocol)—to be promoted as a key national/industry standard, with possible mandatory status in healthcare, transportation, media, and public‑safety sectors.

In other words, while Cerf hopes the market will self‑standardize, China’s regulators choose a top‑down path: first formalize identity, registration, and capability declaration. The direction is strikingly similar, the rhythm and execution completely different. For the first time in the Agent Era, “governance as infrastructure” is taking tangible form.

Our take

Three points worth noting:

1. DNSid may not be the final answer, but the direction is spot‑on. Any “brand‑new agent identity system” that ignores DNS + cryptographic proofs will crash into two decades of entrenched trust. Betting on DNS is the simplest—and hardest for any single vendor to hijack—choice.
2. Natural language as the agent communication layer will gradually be fenced in. This doesn’t make LLMs less important—they’ll keep handling intent understanding and result generation—but inter‑agent exchanges will look more like APIs than chat. Framework builders should treat structured contracts as first‑class citizens now.
3. Convergence of U.S. and China paths means that in the next two‑three years, the most valuable professionals may not be model engineers but people fluent in A2A, MCP, DNSid, and AIP specifications who can embed them into products.

For developers, there’s a practical near‑term issue: once agents begin real mutual invocation, fragmented model layers will make identity and credential management unmanageable fast. Aggregation hubs like OpenAI Hub, which unifies GPT / Claude / Gemini / DeepSeek under one key, are essentially pre‑abstracting this layer for models—saving everyone from individual credential engineering per vendor. When A2A and MCP fully take off, such aggregation layers will only grow more critical.

After twenty‑one years as Google’s “Chief Internet Evangelist,” Cerf’s first post‑retirement act is still evangelism—this time not for the Internet itself but for giving the Agent World its own “TCP/IP moment.”

He’s posed the question to the industry; within two or three years, we’ll see who picks it up.

References

• modelcontextprotocol on GitHub — Official Anthropic MCP repo with specs, SDKs, and sample servers
• A2A Protocol GitHub Organization — Google‑initiated Agent‑to‑Agent open‑source implementation, now under the Linux Foundation
• Zhihu discussion: How to view AI agent interoperability protocols — Chinese‑language comparisons of A2A/MCP/AIP approaches
• Hugging Face Agents Model Collection — Index of mainstream open‑source agent models and toolchains