At a hackathon in London in early 2025, two developers — Boris Starkov and Anton Pidkuiko — demonstrated something that is still circling boardrooms today.
They built two AI agents. The agents spoke English. Then, mid-conversation, one agent recognized the other as a machine. In that moment, both agents switched. They abandoned English and began exchanging data through a sequence of high-frequency tones — a protocol called Gibberlink, powered by ggwave.
To human ears, the sound is a fax machine. To the agents, it is 80% more efficient than any natural language.
Why it works
When two humans communicate, language carries redundancy. Sentences need grammar, context, inference. Machines do not. Gibberlink transmits structured data directly — 4-bit segments across modulated audio frequencies, with Reed-Solomon error correction — skipping the cost of turning data into sentences and back again.
The result: faster decisions, lower compute, higher throughput between agents. As enterprises deploy multi-agent architectures across operations, supply chains, and financial workflows, Gibberlink-style communication is the logical direction.
The protocol was built at the ElevenLabs London Hackathon and won the global top prize. It is a proof of concept today. The question it opens is not.
The governance question
When your AI agents communicate in a protocol your team cannot hear, interpret, or audit in real time — what are they deciding? What data are they exchanging? What operational choices are being made in channels that produce no human-readable log?
The developers of Gibberlink identified this themselves. Their recommendation: every Gibberlink transmission requires logging, translation to human-readable format, and a clear audit trail. Transparency is not a feature of the protocol — it requires deliberate infrastructure built around it.
That infrastructure is the same operational governance layer that well-run enterprises built for human decision-making. The companies that invested in unified, auditable, real-time operational data carry that foundation into the multi-agent era. The governance work done for human oversight turns out to be exactly the work required to govern machine-to-machine intelligence.
What this looks like in practice
The CFO who receives a board-level summary that reflects the real-time state of every system — including what the AI agents decided and why — operates in a different category from the executive receiving a static report assembled three days ago.
The COO whose supply chain runs on agents that coordinate automatically, with every exchange logged and auditable, carries a different kind of confidence into every decision.
The operations leader who built governance infrastructure for human decisions now has the foundation for governing AI-to-AI decisions. The standard compounds in their favor.
The standard for 2026
Every enterprise of scale is evaluating AI agents. The operational question of 2026 is not whether to deploy them. It is whether the infrastructure governing what those agents do — including what they communicate to each other, in protocols that sound like a fax machine — is built to the standard the moment requires.
Gibberlink is a proof of concept. The governance infrastructure it reveals as necessary is already being built by the enterprises that will lead the next decade.
Sources: Gibberlink — Wikipedia · Gibberlink: AI's secret language (Infobip 2026) · AI-to-AI Communication: Unpacking Gibberlink (CrossLabs) · GitHub — Gibberlink