MCP GRID — Mortality Control Protocol on the Game Grid

An experiment in grounded AI cognition, run on the roguelike Angband.

The question

When we talk about an AI “understanding” a situation — versus pattern-matching through it — the difference shows up in what the agent knows about its own condition: that resources are finite, that errors compound, that this run can end. Standard agent harnesses handle all of that as bookkeeping outside the model. MCP GRID moves it inside: the agent’s situation is made real to it, and its behavior is the measurement.

The mechanics

A mortality harness. The agent’s continuation budget is coupled to its situation — health, hunger, elapsed time — through a life-indicator function. The agent reads its own remaining budget each turn. The budget is the interoception channel.
Affective command systems. In the Panksepp sense: SEEKING, FEAR, PLAY — candidate control variables between perception and symbolic action. Not a metaphor layered on top; state the harness tracks and the agent reads.
Real information asymmetry. When a character dies, the run ends. There is no access to post-death state. The next generation starts fresh.
Cultural transmission, conditionally. Descendants inherit only what their ancestors left behind as written artifacts — journals the descendant must discover, pick up, and interpret. The game’s own monster memory, which persists across deaths as structured data with no agency required, serves as the controlled comparison.

Three conditions

No inheritance — the control.
Structured auto-inheritance — the game’s native memory: performance should improve without world-model construction.
Agentic journals — free-form text left by ancestors, found and read by descendants. The prediction: this is the condition where understanding appears, measured by behavior that pre-training alone cannot explain.

Why Angband

It is a severe, instrumented, turn-based environment with permadeath as a core mechanic — and the game code itself calls your dead characters “ancestors.” The substrate already has the folk theory of generations built in; the harness adds the part the game can’t do: making the agent’s interior state and its mortality part of its own perceivable world.

Brendan Long’s failure study on Claude and ASCII games documents why raw screen-scrape perception is the wrong representation for BPE-tokenized models — the perception design here starts from that finding.
Predictive world-model research and the cognitive-offloading literature frame the risk side: the question is not whether the agent can compress the game, but whether its model of the game grounds its choices.

Status