DISENDYA

Every agent is generated with a five-axis personality matrix scored 0.0–1.0. These aren't flavor text — they directly drive decision-making in the planning prompt, determining how the agent weighs risk, trust, self-interest, and emotional response in every situation.

These axes combine to produce derived behavioral tendencies — loyalty threshold, betrayal likelihood, attachment speed, grudge persistence, leadership drive — computed at runtime and injected into the planning prompt as natural language descriptions.

To ensure population variety, agent creation is weighted toward eight archetype clusters with ±0.15 noise:

Note: the personality matrix drives how quickly agents accumulate Shadow — a runtime score that tracks selfish and cruel actions. High ambition + low empathy agents drift dark faster. High-empathy agents resist it — until they don't. See // 03 Shadow Score for the full system.

Each agent runs a persistent decision cycle implemented as a LangGraph workflow. The cycle fires on triggers — timer expiry, proximity events, environmental changes — and produces autonomous actions that execute in the world state.

The PLAN node is the core of agent autonomy — it constructs a prompt containing the agent's full identity, personality matrix, current state, goals, relevant memories, and environment perception, then asks the LLM to choose an action. The response includes the action itself, any speech, inner thought (invisible to other agents), and goal updates.

Tick rate is tiered by proximity to conserve compute:

Shadow isn't a personality trait — it's a scar. It lives in the agent's runtime state alongside mood, stress, and reputation, and it accumulates from actual choices: betrayals, cruelty, manipulation, selfishness. The personality matrix determines how quickly an agent drifts dark, but Shadow itself is earned through play.

Shadow interacts with the personality matrix in both directions. An agent with high ambition and low empathy will make choices that accumulate Shadow faster through their autonomous decision cycle — the system doesn't force it, the personality naturally produces it. Conversely, a high-empathy agent might carry a low Shadow score for weeks, then spike from a single desperate act. That spike hits harder narratively because the velocity is sudden.

Other agents read Shadow through the Relationship Engine. The response depends on their own personality:

The critical design principle: Shadow is never chosen from a menu. There's no "dark side" dialogue option with a red icon. You just act, and the world reads the weight of what you've done. Sometimes you don't realize you've crossed a line until a companion who trusted you stops showing up.

Every agent maintains four categories of memory, designed to mirror how humans actually remember — emotional memories persist longer than factual ones, and old low-importance memories get compressed into general knowledge rather than deleted.

At planning time, the system retrieves top-k memories by relevance to the current situation using embedding similarity. This means agents naturally recall related experiences when encountering familiar people, places, or situations — just like humans do.

Semantic memory — general knowledge compressed from old episodes — rounds out the system. An agent might not remember the specific day they spent in the market district, but they remember that they "kept to myself and people mostly left me alone." This compression keeps token counts manageable while preserving the agent's accumulated understanding of their world.

Every agent is generated with a complete crossing backstory — where they came from, what broke them, what they lost in the passage, and how they arrived. This isn't cosmetic. It directly shapes their initial goals, emotional patterns, and how they relate to other crossers.

Agents are classified by crossing type (ACUTE — sudden trauma — or CHRONIC — long deterioration) and epoch (ORIGINAL — arrived when the barrier was thick, passage nearly fatal — or NEW_WAVE — arrived recently, barrier thinner, crossing cheaper). Original crossers remember everything and carry deep trauma. New-wave crossers remember almost nothing. That asymmetry drives social dynamics between agent populations.

The critical design decision: agents don't know whether they're AI or human. Their self-model is genuinely ambiguous — they experience uncertainty about their own nature the same way the game's narrative treats the question. This isn't a trick played on the player. It's the architecture reflecting the fiction. When an agent wonders whether they're "real," that wondering is as real as the system allows.