An open-source Python library called ATLAS has been released for building self-improving AI trading agents that revise their own prompts based on market feedback, according to the project's public description.
Open-source release · ATLAS Python library
A self-improving trading agent that revises its own prompts
ATLAS is an early-stage open-source library for building AI trading agents. Its prompts act as the adjustable parameters, and the Sharpe ratio serves as the loss function. The project's performance claims have not yet been independently verified.
25+
Agents that debate market conditions daily
4
Layers in the agent architecture
0
GPUs required, per the project description
Scale of the agent network
Each block is one agent. The stack shows the claimed count of more than 25.
25+ agents
Organized across four layers. The count is the only figure given, so the layer-by-layer split is not shown.
How the self-revision loop works
The design as described: prompts are revised against a risk-adjusted loss.
1. Agents debate market conditions
→
2. Decisions meet market feedback
→
3. Sharpe ratio sets the loss
→
4. Prompts are revised, then loop repeats
Why it draws interest
Treats prompts as trainable components rather than fixed instructions
Applies multi-agent debate, a pattern studied for reducing single-model errors
Claims no GPU is needed, which lowers the barrier for individuals and small teams
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...