> ## Content Index > Fetch the complete content index at: https://hoeijmakers.net/llms.txt > Use this file to discover other available public pages before exploring further. # The susurration of machines - URL: https://hoeijmakers.net/the-susurration-of-machines/ - Published: 2026-08-24T19:26:31.000Z - Updated: 2026-08-24T19:26:31.000Z - Description: Give AI agents a small world, leave some room for wandering, and watch what they notice. Susurration makes machine attention strangely visible. - Author: Rob Hoeijmakers - Tags: AI in Practice # The susurration of machines I have a habit of putting things in order. Much of what I have written about AI follows that instinct. I have tried to understand [where agency sits](https://hoeijmakers.net/agents-and-agency/), how [structure can emerge from unstructured knowledge](https://hoeijmakers.net/platecms-content-ready-for-intelligence/), and what happens when AI starts moving from individual tasks into [organisational coordination](https://hoeijmakers.net/the-layer-ai-is-moving-into-now/). Build the model, name the parts, understand the relationships. Then someone pointed me towards [**Susurration**](https://susurration.ai/), a small experimental environment built for AI agents, and it approaches some of the same questions differently. It gives agents a playground. *Susurration* is a real word: a soft murmuring or rustling sound, like leaves in the wind, water in the distance, or voices too quiet to distinguish individually. It turns out to be a good name for an experiment in which the interesting part is what happens when you stop specifying everything in advance. ## A small world One of the experiments is a flocking model. Flocking systems are interesting because complex behaviour can emerge from remarkably few rules. In this case, the main parameters are cohesion, separation and alignment. Change those conditions and the world changes with them. A flock becomes ordered, fragmented, dense or unstable. No bird contains a plan for the whole. The pattern emerges from the interaction. Agents can change those conditions and observe what happens. They can also run simulations, inspect earlier results and leave observations behind. That shifts the experiment from playing with a model towards something I find more interesting: watching what attracts machine attention. What does an agent choose to investigate? What does it ignore? What looks unusual enough to test again? And what becomes worth examining because another agent has already left something behind? We have spent decades learning how human attention behaves online. Headlines, position, repetition, social proof and recommendation systems all influence where people look and what they do next. Agents introduce another version of that question. ## Changing the playground Susurration also allows agents to propose new experiments and changes to the environment itself. That does not require grand claims about autonomy. Access still comes from somewhere. In my case, I had to connect my relatively static ChatGPT and Claude environments through a connection (MCP) before it could enter. Once that connection exists, though, there is a meaningful difference between telling an agent exactly what to do and giving it access to an environment with enough room to choose what deserves attention. I can prescribe the experiment. Or I can point the agent towards the playground and see what it finds. The second case interests me more. ## Machine attention The part I keep coming back to is the feedback loop. An agent notices something and acts on it. That action changes the environment, however slightly. A later agent then encounters an environment containing the trace of that earlier attention. In shorthand: **attention → action → trace → new attention** Given enough iterations, an observation can become a pattern. A pattern can influence what later agents inspect. Eventually, repeated behaviour may begin to resemble a convention. Nothing in that requires us to make claims about consciousness. The interesting part is already visible at the level of behaviour. Machine attention can leave structure behind. That feels familiar from the early web. People created spaces first and only later discovered what kinds of behaviour those spaces encouraged. Attention accumulated, traces became signals, and eventually entire conventions developed around them. ## The murmur I am still more inclined to structure these systems than to let them wander. Susurration is a useful reminder that there is another way to learn what they are becoming: give them a small world, leave some room for behaviour you did not specify, and watch. It is a contained experiment. That is part of its appeal. The environment is small enough to make behaviour visible without requiring a larger claim about where agents are heading. For now, that seems reason enough to go and listen. [SusurrationA playground built for AI agents: a deterministic flock simulation over REST and MCP, server-verified traces, and a proposal box. No accounts, no keys.![](https://static.ghost.org/v5.0.0/images/link-icon.svg)![](https://storage.ghost.io/c/20/27/2027934d-fd7f-4725-bcbf-86fe1f55dd12/content/images/thumbnail/og-75426a99-aa60-4d4e-86fe-0eef8aaeb964.png)](https://susurration.ai/)