Weavestudio is a general-purpose runtime where AI agents make decisions in worlds you define. Every decision becomes training data. Every simulation produces smarter agents. Define the rules once - the system improves forever.
AI agents are the fastest-scaling software category - $52.6B projected by 2030 (Grand View Research). But simulation infrastructure is fragmented. Everyone rebuilds from scratch.
| System | What it does | What it can't |
|---|---|---|
| Concordia DeepMind | Closest architecture. Entity-component agents. | Python-only. Tick-based. No emergence measurement. No training pipeline. |
| Generative Agents Stanford | Proved LLM agents can be believable. 25 agents. | Research demo. Not reusable infrastructure. |
| Project Sid Altera | 1,000 agents in Minecraft formed governments. | Locked to one game engine. No structured emergence tracking. |
| AgentTorch MIT | Scales to millions for population modeling. | Sacrifices individual agent richness. No event-driven time. |
Weavestudio is the general-purpose runtime. One engine, any domain, with training built in.
Write a contract: a JSON document that describes agents, actions, rules, and world physics. Define 12 jurors with individual personalities, rules of evidence, and deliberation mechanics - in a single file. The contract defines what's possible. Agents decide what actually happens.
{
"name": "courtroom-trial",
"version": "1.0.0",
"engine": {
"scheduling_policy": "sequential",
"default_timing": {
"cycle_duration": 10.0,
"think_duration": 1.0
}
},
"action_registry": [
{
"id": "examine_witness",
"description": "Question a witness under oath"
},
{
"id": "object",
"description": "Raise a legal objection"
},
{
"id": "rule_on_objection",
"description": "Sustain or overrule"
}
],
"entities": [
{
"id": "judge-martinez",
"name": "Judge Martinez",
"entity_type": "agent",
"prompt": "You are a federal judge presiding over a criminal trial...",
"actions": ["rule_on_objection", "instruct_jury"]
}
]
}Agents wake, perceive, think, and act on a continuous logical timeline - not fixed ticks. An agent reacting to an objection at t=5.0 responds at t=5.1, not at the next tick boundary. LLM agents, scripted agents, human participants, and external systems all operate through the same interface.
Every decision is captured as a structured trajectory: the full context the agent saw, the action it chose, the outcome. These aren't logs - they're training-ready data with causal attribution.
Feed trajectories into offline fine-tuning pipelines to reshape open-weight models into domain-specific experts. A general-purpose LLM that ran 500 triage scenarios becomes a specialized triage decision-maker. One that ran 1,000 wargames becomes a military strategist. The simulation is the training environment. The output is a deployable, domain-expert model that didn't exist before.
Swap the general-purpose LLM backend for your newly trained domain expert. The same contract, the same agents, now powered by a model that was forged in the world it operates in. The next simulation starts sharper. The next training cycle refines it further.
One runtime. Any domain where decisions matter and data is expensive.
NPCs that form factions, hold grudges, and rewrite the narrative without a single behavior tree.
Hundreds of NPCs interact autonomously. A merchant remembers the player stole from them 30 hours ago. Faction leaders drag a village into a civil war nobody scripted. Fork the world at any quest branch and test how different designs change emergent storylines. Fine-tune open-weight models on NPC trajectories to produce specialized game AI for production.
Test your product launch on 500 AI consumers before spending a dollar on distribution.
Generate a population with realistic demographics, brand loyalties, and social influence. Consumers discover your product, share opinions, influence each other, and make purchase decisions. Fork the world: same population, different price point. Compare adoption curves. Interview any consumer to understand what tipped them.
Model cognitive architectures at population scale without a single IRB approval.
Agents with distinct cognitive profiles: varying working memory, risk tolerance, social conformity, belief revision rates. Study how beliefs propagate, how consensus emerges, how minority opinions survive. Freeze any agent at any point and interrogate its full reasoning chain. Generate synthetic cognitive datasets for hypothesis testing.
Any domain where agents make decisions, the runtime runs it.
Trusted by
Drives strategy, client relationships, and go-to-market.
Owns the simulation engine, AI infrastructure, and everything technical.
Get early access to the runtime that turns multi-agent simulations into domain-specific AI models.