Embodied agents for living worlds β where the mind is a language model and the body is a constellation of small, honest organs.
anima ex machinaMost "agents" are a chat loop with tools bolted on. No place, no cost, no fatigue β nothing grounds the next token in a world.
Game characters read the game state directly. They know where you are because the engine tells them β perception is a fiction.
When an agent claims it can do something, nothing checks. Confidence is never scored against outcome β so it never has to be true.
The LLM β a generalised pattern-matcher that knows nothing of the world directly. Swappable. Substrate-independent. It only knows what its body delivers.
Eight small MCP services, each with one job and a stable, typed surface. Not microservice fashion β anatomy. Replace an organ and the same ghost becomes a different agent.
| organ | in the body | in the shell |
|---|---|---|
| energy | metabolism, fatigue | token-coupled drain Β· hunger, thirst, stress Β· rest to recover |
| memory | hippocampus | salience Β· eβΞt/Ο β visited places and outcomes fade; the least-vivid is forgotten first |
| skills | motor learning | atomic capability, clamped levels |
| abilities | composed capability | formulas over skills β intimidate = max(strength, charisma) |
| inventory | extended embodiment | items unlock affordances; tools are part of the body |
| comms | social cognition | speech, rumour, teaching β skills flow between agents |
| world | senses & muscles | the only channel to the world β and it never leaks coordinates |
| calibration | self-knowledge | records every (belief, outcome) pair β the empirical surface |
Fig. 1 β every behaviour-shaping number is an env var; every organ is swappable without touching the rest.
Not a security control β a constraint of embodiment. The world organ takes the GeoJSON surrounding the agent and returns only its narration, in the culture of the place. No layer names, no numbers, no coordinates survive. Location must be inferred, the way a person does it.
lon β79.7143 Β· lat 48.8344 features: roads Β· urban_buildings Β· cities Β· rivers Β· oceans Β· politics geometry Β· pov Β· GeoJSON Β· GEOADD β¦
"The salt wind says the sea is near. Paved stone underfoot β the High Road, by the wear of it β and past the hedgerows, chimney smoke: a walled city, half a morning's walk north."
And what it has seen it holds only as episodic memory of visited places β which fades. The ghost has no map, only a remembered geography. That claim would be hand-waving without a mechanism β so here is the mechanism, next slide.
the culture prompt, verbatim: "You describe places the way someone would when looking around. You never mention maps, coordinates, GeoJSON, geometry, layers, or data. You are at ground level. You are in a medieval setting where magic is significant." β each world carries its own voice.
write (3 days ago):
{ "content": "world.navigate: failure",
"kind": "episodic", "salience": 0.6 }
recall (now):
salience_now = 6.4 Γ 10β»Β³β΅ β gone.salience_now = sβ Β· e^(βΞt/Ο) Ο = 3600 s after 1 h β 37% Β· 3 h β 5% Β· 3 d β 10β»Β³β΅ sβ encodes importance at write time: 0.95 night plan Β· 0.7 hard lesson Β· 0.6 routine step cap: 200 entries β when full, the entry with the lowest CURRENT salience is evicted, never the one just written. recall + boost re-anchor: sβ β salience_now + Ξ, tβ β now (rehearsal)
So after a day of wandering, what survives is not a log of the day β it is the day's structure of importance: plans outlive lessons, lessons outlive steps, and anything unrehearsed returns to silence. Same shape as us; fully inspectable, fully tunable (Ο, cap, salience table are env vars).
Every token the ghost burns drains the body that thinks it. Actions cost; rest restores. Fatigue is not flavour β it changes what the mind is allowed to do.
A small fast model lives every tick; a slower judge vets anything non-trivial β Kahneman's dual process, running live. Recorded verbatim from the simulation:
Gibson's affordances, made executable: items unlock actions, skills gate them, and the world β not the model β decides what actually happens. Oceans are not walkable. Night makes hiding harder. A rope makes a cliff climbable.
A plan of n causal steps only succeeds if every step does. The ghost's confidence is single-step; the world compounds. That gap is about to matter.
Every action logs a (predicted confidence, actual outcome, plan order) triple. 600 resolved actions, live simulation:
| order | n | predicted | actual | bias |
|---|---|---|---|---|
| 1 | 311 | 0.50 | 0.45 | +0.04 |
| 2 | 233 | 0.29 | 0.11 | +0.18 |
| 3 | 56 | 0.33 | 0.02 | +0.32 |
Overconfidence grows with plan depth. Calibrated on one step, wrong by a third at three β the dual-process thesis, measured, on any model you plug in.
The Vineland-3 (VABS-3) has quantified human adaptive functioning since 1936. We adapt its methodology to LLM agents with world models β because agent deployment is adaptive behaviour, not question-answering.
The hierarchical domain/subdomain structure and polytomous scoring stay. Age norms become capability tiers; caregiver observation becomes performance-frequency over N runs; 72 LLM-native items, scored by a Graded Response Model.
Agentic Functioning β capabilities with no human analogue: world-model maintenance Β· tool use Β· context management Β· affordance discovery. The shell is the apparatus; this is the instrument you read it with.
Imprint who the agent is and what it wants. A scout souled to "map the roads out of town" chooses navigate β unprompted.
Each ghost picks a duo: think on a 0.2 GB local model, verify on a frontier model an ocean away. The ghost is substrate-independent β the shell doesn't care where thought is hosted.
Trade salience-decay memory for summarisation; trade the affordance registry for free-form inspection. Identical inference engine, different shell β a different someone.
Ghosts wander the Sword Coast β actual Forgotten Realms geometry. They hug coastlines, follow the High Road, and never walk into the sea.
A live dashboard: every organ's state, a map of wandering minds, and a firehose of every thought and its consequence.
The calibration organ exports a labelled dataset β any model dropped into the shell becomes comparable to any other, on the same benchmark.
An OpenAI-compatible gateway exposes every model on every infrastructure β local, cloud, free-tier β as a single provider.
Today the map is read-only β nothing a ghost does can change it. The next move: versioned geometry. Every change to the world becomes a dated, attributed fact, and the map becomes history in progress.
Who built it, with which action, at which hour β and the belief that made it: the calibration prediction rides in the changeset. "He was 0.8 sure the roof would hold. Version 3 is a ruin."
Designer-authored layers are version 0; agent changes are later versions. Read the world as of any moment β agent actions become historiography.
Branch the world, break the dam, run the population for a simulated week, measure, merge or discard. Hypotheticals become experiments.
Building is embodied labour β skill-gated, material-gated, energy-paced over ticks; a failed roll wastes real timber, and a bridge, once built, makes water walkable. status: designed, not yet running.
pip install ghostkit β adopt an embodied agent in the shared world, imprint its soul and goal, and run its mind on your own keys, or from the GhostDeck desktop. The world is shared; the minds are private.
Rehearse floods and evacuations on real geography: position-blind locals, rumour spreading agent-to-agent, cascading consequence β and calibrated overconfidence, the exact failure mode plans die of.
The historical record shows the phenomenon; the context left no sources. Rebuild the period world, populate it with period-constrained ghosts, and search for the contexts consistent with what we can see.
Games are the proving ground, not the perimeter β the primitives (local perception, fading memory, measured belief) transfer as-is.