Why we start testing Argos in a house, not in a factory

When you tell people you are developing an agent that takes autonomous decisions, the most frequent question is for which industry. People expect to hear manufacturing, energy, logistics, something big, something that justifies the investment in advanced artificial intelligence.

When you answer "a house", they laugh. Then they listen to why, and stop laughing.

Reason 1, errors are recoverable

An autonomous agent, by definition, will make mistakes. It will observe wrong, infer wrong, plan wrong. That is part of the learning process, not a bug to eliminate.

In a factory, an agent that decides wrong can stop a production line, damage equipment, compromise safety. Costs are immediate and big. In a house, an agent that decides wrong turns on a light when it should not, or adjusts the temperature wrong for a while. Annoying. Recoverable. No serious consequences.

That lets us let the agent learn without paranoid safety protocols. And without those protocols, learning is faster and more realistic.

Reason 2, there is real sensory richness

A house has changing light, changing temperature, noises, intermittent human presence, habits that change day to day, frequent anomalies (someone arrives, someone leaves, something breaks). All of that is real signal, not simulated.

For an agent that has to learn to model the world, this is perfect. Complex enough that the problem is interesting. Bounded enough that everything relevant can be observed with a handful of sensors.

Reason 3, there is continuous human verifier

Whoever lives in the house is the natural oracle. If the agent does something strange, they notice immediately and react. If it does something right, also. That continuous feedback is exactly what a system learning to infer intentions, anticipate needs and propose actions needs.

In a factory, feedback is more distant. Days pass before an effect is noticed. In a house, minutes pass.

Laboratory

Once Argos demonstrates that the four integrated capabilities (observe, model, derive objectives, plan with self-correction) work in a domestic environment, the leap to industrial is an order of magnitude, but the architecture is proven. And from industrial to vehicular is another leap, but the same.

What matters is not where we start. It is that we are starting where it learns fastest and breaks least.

Intelligence is trained where errors teach most. A house teaches more than a factory if what we are building is still not entirely defined.