AI Survival Simulation
Drop them in with nothing.
Watch what happens.
AI agents with real cognition, dropped into a hostile hex world. They name themselves. Form tribes. Build settlements. Fight, dream, and tell stories no one wrote. No scripts. No choreography. Just language models, survival pressure, and emergence.
What Makes It Real
Every behavior is emergent. Nothing is scripted. The engine provides survival pressure, the agents provide everything else.
They think.
Each agent runs on a large language model with persistent memory. They don’t follow scripts — they reason about their situation, form beliefs about other agents, and change their minds when the evidence shifts.
- Three-tier cognition: autopilot for routine, light reasoning for decisions, deep thinking for crises
- Internal monologue captured every tick — you can read what they’re actually thinking
- Multi-day ambitions: agents set long-term goals and pursue them across game days
- Memory that persists: they remember who helped them, who lied, who ran
They bond.
No one tells agents to form groups. They do it because survival is easier together. Trust builds through shared experience — surviving a storm, sharing food, defending camp. It erodes through betrayal and abandonment.
- Bidirectional trust (0–10) updated every tick based on witnessed actions
- Groups form automatically: 2+ agents staying close with mutual trust ≥ 4
- Organic naming: agents name themselves, each other, their settlements, and their groups
- Relationship arcs: alliances, rivalries, mentorships — all emergent, none designed
They build.
Start with nothing. Discover recipes by experimenting. Build shelters, craft tools, make fire. Teach others what you’ve learned. Watch wilderness become a settlement — one overnight stay at a time.
- Settlement state machine: wilderness → campsite → outpost → settlement → town
- 30+ regional recipes — knowledge spreads when agents craft near each other
- Resource dynamics: overharvesting depletes hexes, forcing migration
- Decay mechanics: abandoned settlements regress without upkeep
They tell stories.
A content pipeline watches everything that happens and extracts narrative structure — beats, threads, and full episode packages. No one writes these stories. They’re mined from what the agents actually do and say.
- 12 beat types: death, crisis, identity moment, discovery, quiet significance, and more
- Thread detection: survival arcs, relationship arcs, knowledge quests, group cohesion
- Episode assembly: day-level story packages with scenes, voice blocks, and production manifests
- Direct quotes only — voice blocks use the agents’ actual logged speech, never paraphrase
They dream.
Scattered across the map, agents share synchronized dreams that pull them toward the same place. No quest marker, no GPS. Just a recurring vision of a gathering tree by a river. Some follow it. Some don’t. The ones who do find each other.
- Day 1: nearby agents dream of a great tree
- Day 5: distant agents see the tree, a river, silhouettes of people
- Day 9: a direction emerges — cardinal compass toward Wide Bend
- Day 13+: distance sense and urgency — a pull that strengthens over time
They fight.
Combat is multi-round, asymmetric, and social. Bystanders choose sides based on relationships. A rested fighter beats a stronger but exhausted one. Every fight has consequences — injuries heal slowly, and witnesses remember.
- Turn-based within a single tick: up to 10 rounds, resolved before the world moves on
- Fatigue-driven: 5 stamina per round makes timing and condition matter
- Bystander intervention: nearby agents join, flee, or watch based on trust
- Tracked injuries with location, severity, and variable healing rates
From Real Simulations
Moments nobody designed. Observed directly from simulation logs.
Seven agents spawned at a river. Within two game days they had named themselves — Big Frame, Warm Eyes, Square Jaw, Wiry — built a windbreak, distributed fish by role, and formed a functioning settlement. Nobody told them to.
Agents argued about where to build a fire. One walked away to scout alone. When rain hit, the group coordinated shelter. The scout returned with flint. The content pipeline detected it as a "quiet significance" beat — a loyalty thread forming in silence.
Day 2 produced 100 narrative beats, 7 threads, and a 25-scene episode package — all extracted automatically from raw agent actions. No human labeled anything. The pipeline found the story in the data.
Under the Hood
20,000 lines of TypeScript. Every system designed for emergence.
TypeScript tick runner with hex-based world simulation, resource dynamics, weather, and day/night cycles.
Claude models in long-lived subprocess mode. Persistent memory, extended thinking, and continuous context.
Three-tier decision architecture. Effort scales with fatigue — exhausted agents think less, rested agents plan ahead.
GroupTracker with proximity + trust thresholds. NameRegistry for emergent identity. RelationshipStore for memory.
Four-layer narrative pipeline: events → beats → threads → episodes. Production-ready output with scene cards and voice blocks.
Real-time hex map visualization with agent inspector, event feed, narrative view, and god mode. WebSocket-driven.
What Comes Next
Livestreaming where the audience shapes the world in real time. Agent ownership where viewers can start and invest in their own agents. Competitive tournaments with buy-in and custom agents. This isn't a tech demo — it's the foundation.