Gibberlink: AI’s secret language and what it means in 2026

In March 2025, a short video went viral. Two AI agents are on a phone call. Midway through, one asks: "Before we continue, would you like to switch to Gibberlink mode for more efficient communication?" Seconds later, the conversation dissolves into beeps and chirps, machine sounds completely unintelligible to the human ear.

The clip hit 15 million views on X in a week. People were fascinated, unsettled, and had a lot of questions.

Fifteen months on, the picture is clearer. Gibberlink is still a proof-of-concept, not a production standard. But the question it raised has become a real standards debate, one that Google, OpenAI, Anthropic, and Microsoft have all committed to answering.

Here’s what Gibberlink is, how it works, and where it actually fits.

What is Gibberlink mode?

Gibberlink was created by developers Boris Starkov and Anton Pidkuiko at the ElevenLabs London Hackathon in early 2025, arguing that AI agents don’t need to speak like humans when talking to each other.

Natural language is built for people. It’s full of redundancy, ambiguity, and context-dependence. When one AI agent needs to hand structured information to another, speaking in full sentences wastes time and compute. Starkov and Pidkuiko built a system that lets agents recognize each other and switch to a faster, machine-native protocol mid-conversation.

That protocol is called GGWave. It transmits data through modulated audio signals using Frequency-Shift Keying (FSK). The result sounds like dial-up modem tones. To a human ear, it’s meaningless. To the receiving AI, it’s precise, structured, and significantly faster than speech.

80% efficiency boost when AIs communicate in Gibberlink, according to the original demo data.

Gibberlink is open-source and available on GitHub.

How does Gibberlink work?

The Gibberlink process has three steps:

• Two AI agents begin a conversation in standard human language.
• One agent detects it’s talking to another AI, not a human, and proposes switching protocols.
• Both agents switch to GGWave, transmitting structured data over modulated sound waves instead of speech.

The protocol switch isn’t hardcoded. The agents aren’t following a rule that says "if AI detected, switch to GGWave." They interpret conversational context, confirm mutual understanding, and coordinate a behavioral change through the conversation itself. The communication layer emerges from agent reasoning, not from explicit routing logic.

GGWave covers a 4.5kHz frequency spectrum divided into 96 equally spaced frequencies. Data splits into 4-bit segments, each transmitted simultaneously across multiple audio tones, with Reed-Solomon error correction for reliability. At 1,200 baud, it transmits roughly 150 bytes per second.

That’s not fast by network standards. The real efficiency gain comes from eliminating a step most people don’t think about. Turning structured data into sentences, and back again, costs compute on both ends. GGWave transmits the data directly.

Where Gibberlink fits in the 2026 AI protocol stack

When Gibberlink went viral, the infrastructure question was wide open. How should AI agents talk to each other at scale?

That question now has answers.

In 2025, two protocols emerged as serious candidates.

MCP (Model Context Protocol) was released by Anthropic in late 2024 and adopted across the industry in 2025, including by OpenAI, Google DeepMind, and Microsoft. MCP standardizes how AI agents connect to tools and external data sources. By May 2026, MCP has 97 million monthly SDK downloads and over 10,000 active servers. It solves agent-to-tool connectivity.

A2A (Agent-to-Agent Protocol) was created by Google in April 2025 to answer a different question. How do agents from different vendors and frameworks collaborate with each other? A2A launched with 50 partner organizations and grew to 150+ by mid-2025. It’s the closest thing AI agents have to HTTP.

Both protocols are now under the Linux Foundation’s Agentic AI Foundation (AAIF), co-founded by OpenAI, Anthropic, Google, Microsoft, AWS, and Block. The protocol debate effectively settled in December 2025, according to Zylos Research.

MCP and A2A run over standard network connections. Gibberlink addresses what happens in voice channels, where APIs aren’t available mid-call. That distinction matters for call center automation, voice assistants, and any deployment involving phone calls.

As companies replace call center functions with AI agents, the question of how those agents coordinate during a live call becomes practical. Gibberlink is one answer. Whether it becomes a standard or gets replaced by something more robust is still open.

What Gibberlink actually gets you

Speed matters less than you might think. The demo claimed 80% faster exchanges and 90% lower computational load, figures from the original hackathon demo, not an independent study. Those numbers reflect the cost of generating and parsing natural language, not raw transmission speed. For high-volume call centers running thousands of concurrent interactions, that overhead compounds across every call. But that’s not the main reason you’d care about this.

The more useful thing is precision. Natural language is ambiguous. "Around 3pm" means something different depending on context. Structured data doesn’t have that problem. When agents transmit via GGWave, the data is exactly what it is. For booking confirmations, order status checks, and payment processing, that matters in practice.

The autonomy angle is worth naming too. Agents that switch communication modes mid-task don’t need a human in the loop at every step. That’s what starts to make fully autonomous agent workflows achievable rather than theoretical.

What this means for businesses

By Q1 2026, 80% of enterprise applications had embedded at least one AI agent. IDC and McKinsey converge on roughly $1.4 trillion in global enterprise AI agent spend by 2027. Twenty-two percent of production deployments now coordinate three or more agents simultaneously. (Digital Applied, AI Agent Adoption 2026)

The question for most enterprises is no longer whether to deploy AI agents. It’s which communication infrastructure those agents run on.

For end users: Transparency over efficiency

Consumers accept AI handling routine interactions. Package tracking, account queries, booking modifications. They get more uncomfortable as the stakes rise. The moment agents switch to a protocol humans can’t interpret, the transparency burden lands entirely on the business.

That means clear disclosure, logging, and the ability to explain what happened in plain language when a customer asks. Not because regulators demand it (though some do). Because that’s what builds trust over time.

For enterprises: Governance before deployment

The developers behind Gibberlink advocate logging all transmissions with human-readable audit trails. The AAIF’s governance work on MCP and A2A points in the same direction. Efficiency gains without accountability structures create more risk than they solve.

Before deploying any agent-to-agent communication layer, you need a logging framework, anomaly detection for unusual patterns, and a defined path to human escalation. In that order.

Where this plays out in practice

Customer service and order management

A customer contacts an eCommerce support agent. Behind the scenes, that agent needs inventory data from a warehouse system, shipping status from a logistics provider, and returns policy from an internal knowledge base, all in seconds. With agent-to-agent communication, those queries happen in parallel. The customer doesn’t wait through each handoff. Gibberlink’s approach applies specifically to voice channels where APIs aren’t available mid-call.

Travel and booking

A virtual travel agent handles a flight rebooking during a disruption. It checks availability across multiple airlines, confirms hotel alternatives, and updates ground transport, all simultaneously, across different providers’ systems. AI-to-AI communication compresses what would be a 15-minute call into near-instant resolution.

Telecom and network operations

For mobile network operators (MNOs) and MVNOs, agent-to-agent communication has direct application in eSIM provisioning, roaming management, and customer onboarding at scale. When a device activates in a new country, automated agents check coverage agreements, provision local data plans, and send activation confirmations without human intervention at any step.

Infobip’s SMS API and messaging infrastructure handle the human-facing side. Activation alerts, roaming tips, plan upgrade prompts, all triggered automatically based on what the agents resolved.

Smart city and IoT coordination

City infrastructure already runs on agent-to-agent logic. Traffic management, waste collection routing, and public transit scheduling all involve dozens of systems exchanging data continuously. Gibberlink’s audio approach is less relevant here since network APIs are available. But the underlying principle, agents negotiating protocol mid-interaction, applies directly to how these systems handle edge cases and exceptions.

Limitations to know before you build

Gibberlink is a prototype. The creators are clear about that. The limitations are real and worth naming before you try to build on it.

At 150 bytes per second, GGWave is not a replacement for APIs. It handles structured metadata and short payloads, not large data transfers.

Audio quality matters, too. Background noise, poor microphone hardware, and latency degrade signal in ways that network protocols don’t face. Real call center environments are not demo conditions.

The protocol switch relies on agents correctly identifying each other through conversational context. That works in a demo. It becomes less reliable when models update, prompts change, or conversation flows vary from what was tested.

And there’s no unified standard yet. MCP and A2A don’t address the audio channel. Gibberlink doesn’t address the API layer. Nobody has built something that covers both.

The future of Gibberlink

Gibberlink was a hackathon demo, and to this day it remains one. But the direction it pointed in (machines communicating in their own optimized language rather than mimicking human speech) is increasingly where the industry is heading.

The most likely near-term path for Gibberlink is as a foundation for research into audio-layer agent communication, rather than a production protocol in its own right. Its open-source release means developers can extend it. Higher baud rates, adaptive error correction, integration with specific voice AI platforms. Whether a production-ready version emerges from the community or gets superseded by something built on similar principles is still an open question.

There are also valid concerns worth naming. As agent-to-agent communication becomes more capable, the risk is that transparency gets deprioritized in favor of efficiency. Protocols humans can’t interpret need deliberate logging and audit design baked in from the start. That’s not a Gibberlink-specific problem, but Gibberlink makes it visible in a way that’s easy to understand.

On the other side, the same efficiency gains that make Gibberlink interesting are what let human agents focus on what machines can’t do well. Judgment, empathy, the kind of context that doesn’t fit in a structured payload. Real-time AI translation is already narrowing language barriers. Voice agents that handle routine calls free up human agents for the conversations that matter.

Gibberlink’s creators said from the start they weren’t commercializing it. That may change, or someone else may take the idea further. Overall, Gibberlink’s impact on human communication will likely be multifaceted and depend on how the technology is developed and implemented. It’s up to us to guide its development.

FAQs