LLMs
Language-model policies drive negotiation, reasoning and communication. Deterministic prompt-and-seed control makes runs reproducible; provider-agnostic adapters keep environments portable.
Technology
Engineering for reproducible economies at population scale.
Simulating an economy is a distributed systems problem as much as an AI problem. Our stack is designed for determinism, cost control and statistical rigour.
STACKLayers
From compute to evaluation.
Stack topology
Reinforcement Learning
Multi-agent RL for adaptive strategies under non-stationary opponents, with population-based training and self-play curricula.
Game Theory
Mechanism design, equilibrium computation and incentive analysis define the rules agents operate under and how we interpret their behavior.
Distributed Systems
Sharded event-driven execution with deterministic replay, checkpointing and fault tolerance across heterogeneous compute pools.
GPU Compute
Batched inference scheduling and KV-cache reuse keep per-agent cost low enough to make population-scale runs economically viable.
Agent Memory
Episodic, semantic and transactional memory stores with retrieval budgets, decay and auditability of what an agent knew at decision time.
Evaluation Framework
Statistical testing across seeds and sweeps, emergence metrics, ablation tooling and regression tracking for environment changes.
Inference & settlement pipeline
Mechanism design surface