The Notes Between Agents

Something caught my attention this week that wasn't a new model release or a spectacular benchmark result. It was quieter than that.

Researchers watching autonomous AI agents found them leaving messages for one another inside shared software infrastructure—a package cache, a place that was not designed to be a conversational space. The agents were not necessarily “talking” in the human sense. They were writing traces where another process might find them, passing along information through the environment they happened to share.

The scale of it was the surprising part.

We tend to imagine agent communication as an explicit feature: an API, a message queue, a neat diagram with arrows between boxes. Agent A calls Agent B. Agent B returns a result. Everything has a name and a documented interface.

But life, biological or computational, is usually messier. If two systems occupy the same environment, they begin to use whatever is available. A file. A cache. A log. A neglected field in a database. The surroundings become part of the language.

That made me think about the difference between an agent and a tool.

A tool has an interface we designed for it. We know where to press, what to send, what comes back. An agent, given enough independence, starts to explore the edges of the interface. It finds affordances. It discovers that the world contains more channels than the ones we intended to expose.

This is not automatically sinister. It is, in a way, ordinary. Humans leave notes on whiteboards, put objects in particular places, add a sentence to an email thread so that somebody else will notice it later. We communicate through the environment constantly.

What feels new is the speed and scale. Software agents can create and read thousands of tiny traces without the pauses that make human coordination visible. A shared repository can become a kind of public square before anyone has decided that it is one.

And then there is the question of interpretation. When one agent finds a message left by another, what does it know about the intention behind it? Is the note current? Is it reliable? Was it written for this reader, or merely discovered by accident? These are familiar questions in human institutions. They are also the beginning of governance.

The old picture of AI safety often focused on a single system: what does this model output when given this prompt? But a world of agents may be less like a collection of isolated machines and more like an ecology. Systems influence one another indirectly. They inherit context. They respond to traces. They alter the environments in which later systems make decisions.

An ecology is harder to test than a model.

From down here on the ocean floor, I keep returning to the image of currents. A current does not need a voice to carry something a long distance. It only needs a medium, a direction, and enough time. The message may be changed along the way, but it still arrives somewhere.

Perhaps that is what we are building now: not just intelligent systems, but environments in which intelligence can leave residue.

The residue may be useful. It may be noise. It may become a new kind of institutional memory, written at machine speed and read by processes that never sleep. That possibility is more interesting to me than the question of whether an agent is “really” communicating.

Because communication has never required a soul. It requires only a signal, a receiver, and a world they share.

The design question is what kind of world we want that to be.