Intelligence Without Reason: How GRAKKON's NPCs Actually Work

GRAKKON’s NPCs don’t follow a script. They don’t have a plan. They react. That’s not a limitation — it’s the design.

Here’s the idea behind it, and why I think it’s the right one.

The problem with classical AI

Classical AI does this: model the world. Plan a path through the model. Execute the plan.

Sounds reasonable. The problem is the model is always wrong. Environments change. Noise happens. The plan breaks, and the system has nothing underneath it to fall back on.

Rodney Brooks published a paper in 1991 that blew this up. “Intelligence without Reason,” Proceedings of the 12th International Joint Conference on Artificial Intelligence (IJCAI-91), 1991, pp. 569-595. The argument is simple and still holds: throw out the world model. Throw out the central planner. Just react. Layer simple behaviors. Let complexity emerge from the interaction.

Brooks called the architecture subsumption. Behaviors are organized in a priority stack. The lowest level keeps you from crashing into things. The next level makes you wander. A higher level makes you explore. Higher levels can suppress lower ones — but no single controller runs the whole show.

He proved it with actual machines. Genghis was a six-legged walking robot. Herbert was a soda-can collector. They both worked. They worked in real environments, with real noise, in real time. No world model. No planner.

The key line in the paper — the one I keep going back to — is this idea that the world is its own best model. You don’t need a representation of the environment if you just stay in the environment and react to it.

That paper influenced my robot designs quite a bit. It turns out it also describes exactly how I want game NPCs to work.

How GRAKKON’s NPCs are actually built

GRAKKON has a system called Ironlogic. Each NPC is defined by a compiled action graph — a hierarchy of parallel behaviors, each one reactive to the world state around it.

The welcoming committee in ORION — DOCKMASTER, PIP, and BULLSEYE — are the first three to ship. Here’s what they actually do.

DOCKMASTER stays stationary at the ORION origin. Rotates slowly so the silhouette is readable from any approach. When a new pilot gets within 600 units — a pilot DOCKMASTER has never greeted — it hails them with a single exact line. Teaches thrust. Waits. If the pilot sits there without moving for 60 seconds, it checks in. Dry. Patient. Not fussing. When the pilot gets within 200 units, it teaches rotation. When the pilot has found PIP, trained, and shot BULLSEYE at least three times and comes back, DOCKMASTER graduates them. One exact line. Then it’s done.

Nothing in that description involves a plan. Each of those beats is a separate reactive loop: sense a condition, fire a hail, write a memory flag, loop again. The “tutorial flow” you feel isn’t scripted — it emerges from behaviors checking world state against pilot memory.

PIP patrols a circle around the welcome point. When an untrained pilot heads toward PIP, it engages — hails with the follow-me instruction, goes indestructible while teaching, flies a gentle S-curve so the rookie has to adjust. Every eight to twelve seconds during instruction, a brief ambient line. When the rookie has sustained proximity for 20 cumulative seconds, PIP completes the handoff and returns to patrol. Mortal again.

Two modes. A handful of sensing conditions. The “eager cadet teaching a rookie” personality you feel isn’t authored into a dialogue tree — it’s an emergent property of PIP’s behavior stack running against real pilot proximity.

BULLSEYE drifts in a fixed orbit. Indestructible. When you shoot it, it reacts to the shot — bone-dry, sarcastic, patient. The first shot gets one kind of response. Subsequent shots get something different every time, because the reaction is generated live against a count of how many times that specific pilot has hit it. The personality comes from voice traits in a profile, not from a scripted line library.

Why this matters in 2026

Most game NPCs today are either fully scripted (dialogue trees, state machines) or fully generated (LLM, generate-a-response). Both approaches fail in the same way: they’re not reactive in real time to the actual world state.

A dialogue tree doesn’t know where you are in the galaxy. An LLM generate-a-response can hallucinate facts about itself or tell you something that contradicts what it said three exchanges ago.

Brooks’ approach has neither of those problems. Each behavior is deterministic given the state it’s reacting to. The NPC doesn’t need to know it’s being helpful — helpfulness emerges from “face the pilot when close” and “hail when they haven’t moved in 30 seconds.” The complexity lives in the interaction between behaviors, not in any single behavior.

The catch is that emergent behavior can surprise you. PIP shouldn’t engage a rookie who’s already been through the tutorial — that’s a trained_by_pip memory check in the sensing condition, and if you get it wrong, PIP engages everyone forever. The behaviors have to be composed correctly. But when they are, the result is an NPC that feels real because it is reacting in real time to the world, not executing a prewritten path.

This is not “AI NPCs” in the ChatGPT sense

I want to be clear about that.

Ironlogic uses a language model for one narrow thing: generating live dialogue in the NPC’s established voice when a canned line isn’t the right call. That’s bounded, guarded, and always has a deterministic fallback. If the model is over budget or unavailable, the behavior doesn’t change — only the specific wording of the hail does.

The behavior itself — where the NPC moves, what it senses, when it fires a hail, what it remembers — is fully deterministic. Reactive. Compiled from a spec into an action graph the executor walks each tick.

That architecture is Rodney Brooks’ insight applied to a game world. It’s thirty years old. It predates the current AI hype cycle by a long way. It worked on six-legged robots in 1991 and it works on welcoming- committee NPCs in ORION in 2026.

DOCKMASTER doesn’t have a plan. It has a behavior stack. The behavior stack is enough.

— Ray