Argos: what comes when the agent decides what to do, not when it follows orders

Today everyone talks about AI "agents". But if you look closely at what is running in production, they are executors. They receive a mission, fulfill it, wait for the next one. Faster than a human, no doubt. But the intelligence is in whoever wrote the prompt, not in the agent.

Argos is a project that starts from the other side. The question is not "how do I make the agent fulfill orders better". The question is "what happens when the agent decides what to do itself".

The missing step

For an agent to decide what to do, it needs four capabilities working together, not separately. It has to observe its environment autonomously, without anyone asking for a report. It has to build itself a model of the world from those observations. It has to derive its own objectives from that model. And it has to plan actions with real-time self-correction when the plan clashes with reality.

Any one of those four things alone is an interesting problem. The four together, connected in a continuous loop, is something else. That is what we are researching.

Why we start with a house

Argos starts in a controlled residential environment, a house with sensors, cameras, home automation. There are three reasons to start there, and none of them is commercial.

First, a house is a bounded environment with real sensory richness. There is changing light, changing temperature, noises, intermittent human presence, habits that change day by day, frequent anomalies (someone arrives, someone leaves, something breaks). All of that is real signal, not simulated.

Second, errors are recoverable. If Argos decides to turn on the heating at the wrong time, nothing serious happens. If you did the same in a refinery, you could lose the plant. The house gives us room for it to learn.

Third, there is someone observing all the time. The person living 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.

Laboratory

From the house, the next step is industrial, a plant or production process where sensors and effects are more complex but logic remains manageable. And then, vehicular, where spatial dimension comes into play hard.

Each leap is an order of magnitude. And each leap requires the four capabilities to be really integrated, not working separately. That is what is hard. And that is what justifies Argos having no commercial deadline, because doing it wrong would be worse than not doing it.

What it has to do with Ágora

A lot and a little. A lot because Argos will eventually benefit from Casandra's deliberative engine, the same cross-verification principles apply when an agent decides. Little because the problem we are solving is structurally different, it is not "executing tasks well", it is "deciding which tasks to execute".

It is a long project. It is worth it.

The agent that follows orders already exists. The agent that decides what to do does not yet. That is what we are building.