What Happens When Two Ai Voice Assistants Have A Conversation

This ElevenLabs blog post, published February 25, 2025, explains GibberLink. Developers Boris Starkov and Anton Pidkuiko built it at the company's London hackathon, and their demo went viral (elevenlabs.io). The idea is that when two AI voice agents end up on a phone call with each other, generating human-like speech is wasteful. Starkov called it a waste of compute, money, time, and environment. Instead, the agents should notice they are both AIs and switch to a format meant for machines.

The mechanism is plain. Two agents built on ElevenLabs' Conversational AI start talking in ordinary speech. Each has a tool it can call to enter Gibberlink mode. When the tool is called, the ElevenLabs voice call ends and ggwave, an open-source library that sends data as sound, takes over, still using the same LLM conversation thread. That produces the dial-up-modem chirping heard in the video. In the demo, one agent was prompted to book a hotel room for a wedding and the other played the hotel's booking system. Both were told to switch if they believed the other was an AI agent, but neither was told that it was. The switch came after the customer agent said it was an agent.

The post is open about how ordinary the underlying technique is. Starkov said they chose ggwave as the most convenient and stable option they could find within a hackathon's time limit, that other mechanisms could do the same, and that dial-up modems have sent information as sound since the 1980s. The new part is having LLM agents recognize each other and trigger the switch through a tool call. The post's '80% more efficient' line comes from an X post by Luke Harries of ElevenLabs, and the article gives no measurement behind it. The code is open source on GitHub, and the post notes that ElevenLabs engineers audited it.