Deliberative vs. Reactive Agents

I want to push back on something before we even start. The lesson frames this as a clean trade-off — speed versus strategy — but I think that framing does a lot of quiet damage.

Oh, it absolutely does. 'Speed versus strategy' is the kind of tidy summary that makes people feel they've understood something when they've actually just memorised a slogan. But go on, I suspect your objection is slightly different from mine.

It implies the two are in competition, as if you're standing at a fork in the road and must choose a lane. But most serious AI systems aren't doing that. They're layering both architectures, often simultaneously.

Which the lesson does eventually acknowledge, buried somewhere after the third diagram, in the manner of a terms-and-conditions footnote. But yes. The subsumption architecture. Hybrid approaches. Brooks demonstrated in the eighties that you could get remarkably sophisticated behaviour by stacking reactive layers, no world model required. The deliberative crowd found that philosophically upsetting.

Why upsetting, specifically?

Because deliberative AI came out of a tradition that believed intelligence was fundamentally about representation — you build a model of the world inside the machine, you reason over it, you produce a plan. Brooks came along with his little robots and said, roughly, 'the world is its own best model,' and they took that rather personally.

Which is actually a profound point, and I think it's the thing students most often miss. A reactive agent isn't a dumb agent. It's an agent that's delegated its 'thinking' to the structure of its environment and its sensors. That's a design choice, not a limitation.

It is, and yet people consistently reach for 'reactive equals simple' as a shorthand. A chess engine is deliberative. A fly avoiding your hand is reactive. Therefore reactive is primitive. This is the kind of reasoning that sounds like logic and isn't.

A fly has been successfully avoiding hands for roughly three hundred million years. The chess engine was invented in the forties and still loses to the fly at practically everything that matters outdoors.

I will allow that. That is genuinely the correct point, stated with more economy than I would have managed.

The lesson's third takeaway is the one I think is actually useful — environment dictates suitability. Because that's the question practitioners have to answer. Not 'which is better' in the abstract, but 'what is the environment actually doing, and how much of it do I need to model?'

The computational cost framing is where students tend to go wrong in practice. They see 'deliberative agents are computationally expensive' and they think that means slow laptops. What it actually means is latency at decision time, which in a fast-moving environment is not merely inefficient — it is actively catastrophic. An autonomous vehicle that deliberates for four seconds before reacting to a pedestrian is not a strategic thinker. It is a hazard.

Although the same vehicle needs a deliberative layer for route planning. Which brings us back to the hybrid point, which the lesson treats as a footnote when it's arguably the entire conclusion.

If I were being uncharitable — and I find it difficult to be otherwise — I'd say the lesson teaches you the vocabulary to sound informed at a dinner party. Which is not worthless. But it is not the same as understanding when you'd actually reach for one approach over the other, and more importantly, why that question rarely has a clean answer.

That's fair. Though I'd argue the vocabulary matters, because without it you can't read any of the literature that gets you to the nuance. You have to start somewhere.

You do. I simply prefer students to know they're at the start, rather than concluding they've arrived. Those are quite different cognitive positions, and only one of them leads anywhere interesting.