A weekend hackathon project that lets AI agents talk on the phone with each other in a robotic language, one that’s incomprehensible to humans, has gone viral on social media over the past week.

The project, called GibberLink, was created by two Meta software engineers during a hackathon competition in London, hosted by ElevenLabs and Andreessen Horowitz.

GibberLink allows an AI agent to recognize when it’s speaking on the phone with another AI agent, the project’s creators, Boris Starkov and Anton Pidkuiko, told TechCrunch in an interview. Once an AI agent realizes it’s talking to another AI agent, GibberLink prompts the agents to switch into a more efficient communication protocol called GGWave.

GGWave is an open-source library of sounds in which each sound represents a small bit of data. This lets computers communicate faster and more efficiently than they can by using human speech. To the human ear, however, GGWave sounds like a series of “beeps” and “boops” – exactly what you’d imagine a computer’s native language sounds like.

While it seems unlikely today that two AI agents would end up on the phone with one another, it’s not impossible to imagine these scenarios arising soon. Companies are increasingly replacing call center employees with AI agents from ElevenLabs, Level AI, Retell AI, and other voice-based AI startups.

At the same time, tech giants such as OpenAI, Google, and Amazon are starting to introduce consumer AI agents capable of handling complex tasks on your behalf. These AI agents may soon be able to call a customer service center for you.

In this potential future, GibberLink could enhance the efficiency of communication between AI agents, provided both sides have the protocol enabled. While AI voice models are pretty good at translating human speech into tokens an AI model can understand, the whole process is very compute intensive – and just unnecessary – if two AI agents are talking to each other. Starkov and Pidkuiko estimate that AI agents communicating via GGWave could reduce computation costs by an order of magnitude or more.

For today though, it’s just a cool project. Starkov and Pidkuiko created a website that you can open on two devices to watch as the AI agents talk to each other in GGWave.

Much like a good sci-fi movie, GibberLink’s demo sparked widespread curiosity – and anxiety – about the future of AI agents. In the week since the London hackathon, a video demonstration of GibberLink has amassed over 15 million views on X, and was even reposted by YouTube’s most followed tech reviewer, Marques Brownlee.

However, Starkov and Pidkuiko emphasize that GibberLink’s underlying technology isn’t new – it dates back to the dial-up internet modems of the 1980s.

Some might recall the distinctive sounds of early computers communicating with modems via household landlines – a process known as the “handshake.” Essentially, this handshake represented data transfers using a robotic language, which is fundamentally similar to what’s happening between AI agents through GibberLink.

Starkov and Pidkuiko also noted that the viral craze around GibberLink has taken on a life of its own. Someone purchased the domain GibberLink.com and is now trying to sell it for $85,000. Others have created a GibberLink memecoin, while a few imposters are selling webinars purportedly teaching “agent-to-agent communications.”

Currently, GibberLink’s creators say they are not commercializing the project, and clarify that it is unrelated to their work at Meta. Instead, Starkov and Pidkuiko have open-sourced GibberLink on GitHub, though they say they may work on some additional tooling related to the project in their free time, and release it the near future.