Ambient intelligence
The idea of computers that fade into the background is 35 years old, and the first place it really works is the codebase.

Most of the AI I use still waits for me. Nothing happens until I give it a task, and when it is done it waits for the next one. The agents mostly start with me asking for something. The people who first imagined smart computers had something else in mind. They wanted computers you would stop noticing, that see what is going on and help without being asked, and the name for that is ambient intelligence.
The idea is not new. In 1991 Mark Weiser, a researcher at Xerox PARC, wrote that the most profound technologies are the ones that disappear. His example was writing, which did not go away. We just stopped seeing it as a technology, and he expected computers to go the same way. The term itself came in 1998 from Eli Zelkha and his team at Palo Alto Ventures, in workshops they ran for Philips, the Dutch electronics company.
A few years later the EU made it the guiding vision of its €3.6 billion research program on information technology for 2002 to 2006. Philips went as far as building a test house and watching the people who lived in it through hidden cameras. Not much came of it, and by 2010 researchers were writing about the gap between what ambient intelligence promised and what it delivered. The sensing was never the hard part. Every step from a sensor reading to a sensible action needed its own model, trained on its own labelled data. What we got in the end was the smart speaker, which listened all the time and did nothing until you said its name.
The codebase becomes the room
Language models filled in the missing step, and the place where I see it most is not a home, it’s building software. In ambient intelligence the room notices what is going on and acts without being asked. With agents, the codebase and the tooling around it become that room. I think of it as the old vision arriving somewhere nobody was looking for it.
The old systems needed one model per job. An agent in a codebase runs on one general model, and that model has never seen this code before. It reads the code, the logs, an error message and a bug report written by someone in a hurry, all different kinds of input, and makes sense of them together. It also works out what the person who filed the report probably meant, which is the kind of guessing the old vision struggled with. None of that was trained for this project.
Then it acts. An ambient agent does not wait for anyone to ask. A failing test, a new bug report, a security alert or an outdated dependency is the thing that happens in the room. The agent notices, starts on it without anyone telling it to, writes the fix, runs the tests and comes to a person only to approve or review the change. Given enough freedom, it also decides what to do next by itself, going with whatever looks most likely to work. Ambient intelligence has reached software.
The same thing is actually happening in rooms with walls, only more slowly. Researchers in Liverpool put cameras into a self-driving chemistry lab and pointed ordinary vision-language models at the feed. With no extra training the models spotted a missing lab coat or someone lying on the floor, and the researchers only had to say in plain words what to look for. What counts as a problem was still decided by people. I come from the life sciences, and it is the same arrangement I now have with my agents.
The eyes on the wall
The picture at the top shows a home with eyes on the wall. Weiser saw them coming in 1991. He warned that sensors in every room could make past dictatorships look mild, and he wanted encryption built in from the start. The EU’s own scenarios asked for an off-switch within reach. In a codebase the question is the same one. An agent that sits there sees the code and its whole history, and sometimes credentials and a way into production, so someone has to decide what it may look at and what it may touch.
The eyes can also be wrong. When I work with an AI directly, I see its mistakes, because I review what it hands back. When an agent works in the background, a wrong change arrives looking as finished as a right one, with passing tests and a tidy description. It is not only software. In one study AI scribes, which listen to a doctor’s appointment and draft the note, made up roughly one sentence in seventy. A made-up note looks just as finished as a correct one.
The result is that I spend less time writing each piece and more time on the whole flow. I decide which events an agent may react to, which checks run before anything reaches me, and where a person still has to look. All of it only saves time if I stop checking everything the agents do. That is exactly where it gets uncomfortable, because from then on their mistakes are made where nobody is looking.
You have to trust what they do. Yes, you can have all the evals in place and they help, but trust is not numbers, it’s a feeling. That is true in a codebase and it will be true in a clinic or a home. The technology is finally good enough to fade into the background. Deciding what it is allowed to see and do stays with us.


