Leave a decision to a single AI and it will occasionally give a plausible but wrong answer with full confidence. So what if you asked several AIs the same question and acted only when they agreed? That question started the Bitcoin trading project SOL LABS built and runs itself.
This post is a development story about system architecture. It is not investment advice, and it does not discuss returns.
1. The architecture at a glance
Four models — ChatGPT, Claude, Gemini and DeepSeek — receive the same data and analyze it independently. An order goes to the exchange API only when three or more reach the same decision.
- Collect data such as prices and technical indicators in a common format
- Send it to all four models at once, without showing them each other's answers
- Each model answers only in a fixed format (buy · sell · wait + reasoning)
- If the consensus rule passes (3 of 4 agree), place the order; otherwise wait
2. Why ask them "separately"?
If models can see each other's answers, they tend to follow the first one — much like a meeting drifting toward the loudest voice. So the input is identical, but judgment is fully independent. These models come from different companies and were trained on different data, so one model's mistake is less likely to repeat in the others.
3. Fixing the answer format
Free-form sentences are hard for a program to read, and the same meaning gets phrased in many ways. So every model fills in only the defined fields.
| Field | Content |
|---|---|
| Decision | One of buy · sell · wait |
| Reasoning | Summary of the indicators and reasons used |
Any answer that doesn't match the format is treated as "wait." When in doubt, not acting is the default.
Want to check this against your own data?AI Readiness Assessment · we review your material and map candidates and orderContact us4. Where this applies to business
Getting consensus from multiple AIs isn't only for trading. It applies wherever a wrong decision is costly.
- Two models flag risky contract clauses separately, and clauses both flag get reviewed first
- When models disagree on classifying a customer inquiry, hand it to a person
- Have another model double-check the numbers in a report
The point isn't adding one more AI — it's finding the moments when opinions split and having a person look. Handle the agreed cases quickly and pass the split ones to people, and you get both speed and safety.
Walk us through your workflow and we'll suggest where AI fits and the expected scope.
We build web apps, mobile apps and business systems, then add and run the AI they need. We build and operate our own meeting notes, contract and company analysis AI.
See our projects