I got tired of launching products into the void and wondering "will anyone actually buy this?" So I built a multi-agent market simulation engine.
Instead of asking one LLM "will my product succeed?", MarketFish creates 128+ AI consumers — each with their own identity, budget, emotions, and biases — and lets them shop across 30 rounds. Their purchase decisions, churn patterns, and social influence reveal what real users would do.
How it works:
Stage 1: Build an ontology from seed data (9 real-time APIs: GitHub, HN, ProductHunt, StackOverflow, Google Trends, EastMoney, World Bank, China retail, 36Kr)
Stage 2: Generate a knowledge graph of entities + pain points
Stage 3: Generate diverse AI agents (students, freelancers, SMB owners, enterprise buyers, competitors, macro factors)
Stage 4: Run 30-round market simulation with cross-domain coupling (emotions, social contagion, FOMO) + economic RL
Stage 5: Teacher-student report generation with multi-perspective analysis
Why 6 LLMs instead of 1?
Different agents need different thinking styles. Consumer agents use DeepSeek (fast), SMB owners use Qwen (structured), teacher critiques use Zhipu (skeptical), student reports use Doubao (analytical). 11 providers total, zero dependency on any single vendor.
A fun bug I just fixed:
I added "temporal activation" based on the OASIS paper — not all agents should be active every round. But I set the activation probability to match a 24-hour human cycle (1% at midnight). Result: 128 agents, 1 active per round, zero purchases. Found it after 3 pipeline runs. Fixed by disabling the 24h mapping for simulation timescales.
Today's prediction results:
Woolly AI (auto-haggle discounts) — winner, survival score 0.871
I'm Not Stupid (senior fraud protection) — runner-up, 0.831
Merge AI (smart home control) — 0.829
4 B2B products all failed (0 purchasers)
Built on 6 papers: Generative Agents (2023), OASIS (2025), TwinMarket (2025), Agent Bazaar (2026), EconSimulacra (2026), SMIF.
Open source (MIT). Would love feedback from anyone who's built simulators or market prediction tools.
Instead of asking one LLM "will my product succeed?", MarketFish creates 128+ AI consumers — each with their own identity, budget, emotions, and biases — and lets them shop across 30 rounds. Their purchase decisions, churn patterns, and social influence reveal what real users would do.
How it works:
Stage 1: Build an ontology from seed data (9 real-time APIs: GitHub, HN, ProductHunt, StackOverflow, Google Trends, EastMoney, World Bank, China retail, 36Kr) Stage 2: Generate a knowledge graph of entities + pain points Stage 3: Generate diverse AI agents (students, freelancers, SMB owners, enterprise buyers, competitors, macro factors) Stage 4: Run 30-round market simulation with cross-domain coupling (emotions, social contagion, FOMO) + economic RL Stage 5: Teacher-student report generation with multi-perspective analysis Why 6 LLMs instead of 1? Different agents need different thinking styles. Consumer agents use DeepSeek (fast), SMB owners use Qwen (structured), teacher critiques use Zhipu (skeptical), student reports use Doubao (analytical). 11 providers total, zero dependency on any single vendor.
A fun bug I just fixed: I added "temporal activation" based on the OASIS paper — not all agents should be active every round. But I set the activation probability to match a 24-hour human cycle (1% at midnight). Result: 128 agents, 1 active per round, zero purchases. Found it after 3 pipeline runs. Fixed by disabling the 24h mapping for simulation timescales.
Today's prediction results:
Woolly AI (auto-haggle discounts) — winner, survival score 0.871 I'm Not Stupid (senior fraud protection) — runner-up, 0.831 Merge AI (smart home control) — 0.829 4 B2B products all failed (0 purchasers) Built on 6 papers: Generative Agents (2023), OASIS (2025), TwinMarket (2025), Agent Bazaar (2026), EconSimulacra (2026), SMIF.
Open source (MIT). Would love feedback from anyone who's built simulators or market prediction tools.