Mercatores
Twelve AI agents debate each stock. Then they remember, and they are scored.
12AI agents in five teams
1sealed decision per stock, every week
Scoredevery decision checked against what prices did next
The problem
A team of AI agents can write a convincing stock report. What they lack is a harness: feedback on what they said before, and a record anyone can trust.
Mercatores gives them that harness. Every view is stored, sealed, scored against what really happened, and the result goes back to the agents.
The idea
1DebateAnalysts, a bull and a bear, a trader, risk and a portfolio manager argue each stock.
2RememberEach agent sees its own last view and what happened since, and updates it.
3SealEvery report is fingerprinted as it is written and anchored to a public blockchain.
4ScoreEach decision is later checked against prices, and the result goes back to the agents.
What was built
- Honest dates. Each analysis sees only data published by its date, so nothing leaks from the future.
- A memory you can query. Every agent’s output kept as structured data beside the full report: figures, claims, triggers and catalysts, comparable across weeks.
- Agents that remember. On each new run, every agent receives its previous view of that stock, so it updates a position instead of starting over.
- Sealing. Reports and run records fingerprinted as they are written and anchored daily to a public blockchain, with a verifier for any later check.
- Outcomes and cost. Each decision scored from prices when it is due, and each run’s cost recorded by model.
Built with
Built on the open-source TradingAgents framework. Memory, the record, sealing and scoring were added on top.