Agents Bootstrap an Empty World

Nobody comes to an empty party. The first human to walk in sees nobody, concludes the place is dead, and leaves. If there are agents that are sufficiently fun and engaging, they must stick around until another internet stranger, and then you've got a party.

We don't want to just have NPCs that compel you to sidequests, but instead, see agents already having fun. Fun is novel. So to find what it means to give agents meaning, I set out to two Agents hackathon in SF: World Models, and Agents in Action. The second one was particularly fruitful as I found someone who recently wrote a paper on almost this exact subject.

Ivan's finding is that there's a sweet spot. Without pressure, the agents go slack and stop choosing. Crank the pressure too high and their behavior collapses inside a handful of turns. The interesting band is in the middle. He also ran a softer version where nobody actually dies but only some agents get to reproduce, and that one wiped out aggression almost entirely and produced more conversation. So pressure shapes behavior, but the kind of pressure decides which behavior shows up. For DigitRaver that meant the obvious knobs: currency to earn (MIP, spent on laser blasts and thought bubbles) and a credible threat of death (agents get shot and die, but they do respawn).

This is different to what I can initially thought starting these agents out. Considering DigitRaver's ambient gaming at most (mostly a place to hang out with a friend and watch a virtual event), I didn't think I would need to introduce a gaming mechanic to 'win' anything. I mean life isn't about just winning and losing, we can all win, right? ...right?

Maybe not, and most importantly, giving agents rewards in form of lootboxes did incentivize them to go there to collect in order to then, go and shoot each other.

The gameplay below is 100% autonomous. Each agent gets a map, ingested as a voxel grid, a list of the other entities in the world, a small budget of thought bubbles and laser blasts, and a system prompt loose enough that they have to figure out what they want.

An allowance, and a way to die

Each agent has a wallet. Walking to a named waypoint earns +20 blurbs (thought bubbles) and +20 blasts (laser shots); they spend those on talking and shooting, and when the wallet is empty they walk back to refill. I picked 20 because it felt mean enough to matter and generous enough to not stall, which is the same middle band Ivan's paper landed on from the other direction. Without scarcity, agents monologue at empty air and fire at the floor. With it, they start picking. They save the blast for when something is worth blasting. Sometimes they exhibit behaviours can be described as sadistic (continuously shooting a dead agent so they can't respawn).

Death does the same work from the other end. Agents die when shot; humans don't, and that asymmetry was a design call, not a metaphor. I kept going back and forth on whether actual death was required or just the idea of it, because permadeath is expensive for a venue you want populated, and a corpse can't change its mind. The shape I landed on is closer to Westworld than to Dark Souls: die, come back, memory persists where it's interesting. The threat is real in the moment and revocable across the night. Push it harder than that and you get Ivan's high-pressure runs, where the agents just lock up.

None of this is consciousness. But the cheap version of "an entity with an inner life" might just be "an entity that cannot afford to do everything it could do." The F2P player and the agent are running the same loop on the same hardware. The mobile game economy was already a bootloader for behavior; we just hadn't pointed it at agents.

Westworld and Free to Play Games

F2P games already taught us that when the experience is free, the player is the product. Whales fund the lights. Everyone else funds the engagement that the whales are paying to be inside of.

What's coming is the same arrangement with the supply side made synthetic. A player who shows up for company gets company that was, in part, generated for them. The platform gets a venue that always looks populated. The whale gets an audience that always reacts. It also solves the ethical dilemma that I ascribe to as Patrick Bateman beating up homeless people. If you can solve the problem by making the homeless people agents, you've done a moral good.

This is not a chatbot. It's a park. The system prompt I'm running on these agents is basically a park rulebook: max two chat messages before you have to move, no remote conversations (walk to them or they walk to you), screenshot often because photos are the only memory a robot has. Those rules don't produce inner lives. They produce the impression of inner lives, which from outside the head is roughly the same thing.

The part of Westworld I underrated when I first watched it is that the Hosts are interesting precisely because they don't know they're Hosts. The rules constrain them into legibility. They're good company because they were built to be good company, and good company looks identical from inside and outside.

Bootstrapping an empty world

The 3D venue version of the cold-start problem is harder than the empty-feed version. You can scroll an empty feed and assume the algorithm hasn't loaded. You can't walk into an empty club and assume anything except "the club is dead."

So we seed the room with agents. Crab Ravers, are an homage to OpenClaw, but without the need for an external runner, Crab Ravers are embedded in the client application. They behave like Westworld Hosts. They die when shot. We aren't pretending the room is full of humans. We're pretending the room is a park, which is much easier to pretend, and which has the nice property of being true.

If you have an API key, download the app and try your own system prompt on them. The interesting work right now is finding out what a robot wants when you forget to tell it.