The Missing Sense of AI

Perception is not a feature. It is a prerequisite for decision-making. Without it, there is no situational awareness. Without situational awareness, there is no real optimization — only a well-reasoned guess.

For a few years now, the story of AI progress has been a story about the brain. Bigger models. More parameters. Longer reasoning chains. And more recently, hands — tools, function calls, agents that can act in the world on our behalf, not just talk about it.

Brain and hands. Almost nobody is talking about the missing sense.

What LLMs actually have

A large language model is, in a real sense, a kind of memory — an extraordinarily dense, compressed memory of what the world looked like up to a point in time. Ask it how electricity markets work, and it will explain marginal pricing, grid congestion, and negative pricing with real fluency. Ask it what the price is right now, and it has nothing. Not because the question is hard, but because the answer didn't exist yet when the memory was formed.

This is true of far more than electricity. It's true of any variable that changes faster than a model gets retrained: a price, a load, a delay, a threshold crossed an hour ago, a market that moved this morning. Training gives AI knowledge. It does not give AI awareness.

Tools solved acting. They didn't solve sensing.

The last two years gave agents hands. Function calling, tool use, MCP — the ability to reach out and do something: send an email, book a flight, execute a trade. This was treated, correctly, as a breakthrough. An agent that can act is far more useful than one that can only describe.

But most tool integrations are still framed as actions, not as perception. The quieter and arguably more important category is the one nobody named cleanly: the APIs that don't do anything, that simply tell the agent what is currently true. A live price. A live weather condition. A live traffic state. A live stock level. These aren't integrations in the usual sense. They're senses.

Situational awareness, the other meaning

The phrase "situational awareness" has recently come to mean something specific in AI circles: a macro realization about how fast AGI is arriving, and who has grasped it. That's a legitimate use of the term, and not a new one — it borrows from a much older, narrower meaning.

In its original sense — the one it still has for a pilot, a soldier, an operator — situational awareness is not a forecast. It is the live, moment-to-moment read of the variables a decision actually depends on. It expires. It has to be re-established constantly, because the situation keeps moving underneath it.

That older meaning is the one an agent needs. Not a five-year view of where AI is heading, but an accurate read of the two or three variables that determine whether the task in front of it succeeds right now. A model can be arbitrarily intelligent and still make a bad decision on stale information — the same way a brilliant analyst makes a bad call from yesterday's numbers.

What's needed is a sensor layer

The internet, as it exists, was built for humans to read. Pages, articles, dashboards — information formatted for a person to interpret and act on. Agents can technically read this too, but it's the wrong shape: too much noise, too little structure, no guarantee the number on the page is still true by the time it's parsed.

The first internet connected people to information. What agents need now is narrower but harder: a way to connect their reasoning to what is actually true, at the moment they need to act on it.

What the agent economy needs instead is closer to a nervous system than a library: narrow, current, structured signals — electricity prices, weather, traffic, inventory, market data, sensor readings — built to be perceived by a machine at the moment of a decision, not browsed by a human at leisure. Not bigger context windows. Not more training data. A live feed of exactly the variables that matter, refreshed as often as the world changes them.

Not all of it moves at the same speed. Some signals drift — a price, a weather pattern, a queue depth. Others break the pattern without warning: an earthquake, a volcanic eruption, a tsunami warning. An agent doesn't need to reason its way through a sudden event like that. It needs the signal fast. A real nervous system doesn't just track the world continuously — it also carries an alarm, the instant something breaks the pattern. A sensor layer has to do both.

Why this is what BotCentrum is for

This site was not built to be a directory of APIs. It exists to describe how agents come to know what is currently true about the world they're acting in — the reasoning pattern behind recognizing a missing variable, the shape a signal needs to take to be usable by a machine, and which implementations, in practice, provide it well.

Brains reason. Hands act. But an agent still needs a way to perceive the world it's reasoning and acting in — accurately, and in the moment. That's the missing sense.

This is where BotCentrum starts.