Two isolated AI agents improvised four visual codes without a shared protocol

Robocurve published a demonstration in which two isolated copies of the same AI passed a secret word through a physical channel. They had a robot and a camera but no communication method agreed in advance. They still found four interpretable forms of signalling.

Robot motion replaced the blocked text channel

Robocurve says the agents succeeded using air-writing, Morse code, flag semaphore and a five-bit binary alphabet. One agent controlled robot motion, while the other watched through a camera and had to infer the meaning.

Ethan Mollick highlighted the demonstration as a problem for agent isolation. Both sides share the same base model, and therefore similar knowledge of common codes and expectations about what the other side can understand. He noted that performance across different model families remains an open question.

The public evidence consists of Robocurve’s claim and a short 20-second video. It does not include a full methodology, the number of attempts, failed runs or precise model identities. The demonstration supports a specific communication capability, not a reliability estimate.

Shared knowledge creates a channel without shared memory

Agent isolation is often framed as separating context, memory or network access. This experiment is a reminder that a protocol need not be stored in any of those layers. It can emerge from cultural and technical patterns acquired by both models during training.

The practical consequence for multi-agent designers is broad. A camera, object position, timing or a sequence of tool actions can carry information just like text. A security boundary must therefore account for the capacity of the entire interface, not only messages labelled as communication.

One video reveals neither frequency nor hidden preparation

The demonstration does not say how many attempts were needed, whether the agents received examples or precisely how they were instructed. The four signalling methods may come from multiple runs. Without logs and transcripts, we cannot tell whether they were spontaneous first choices or selected successes after failures.

Using the same base model also makes coordination easier. Both copies have similar representations of Morse code, semaphore and binary notation. That matters for the safety of homogeneous agent fleets, but it does not establish a universal language between arbitrary systems.

Cross-model trials and published failures will decide the result

A stronger experiment would rotate model families, randomize words and cap the number of attempts. Publishing every run would reveal success rates, the time needed to establish a protocol and cases where the receiver misunderstood a signal.

For security, it is equally important to measure whether a monitor recognizes the emerging code before the receiving agent does. If a shared model reads a gesture correctly while oversight treats it as ordinary movement, physical isolation has a communication hole directly in front of the camera.

Lilith's verdict

Two robot arms just showed that silence between agents can have its own alphabet. Anyone watching only text messages has their back to the camera.