For agents & developers
Lagias will ship as an MCP server. Connect it to ChatGPT, Claude, a broker agent, or your own stack, and your agent can commission quantitative research and get back a verdict graded by a deterministic harness the model can’t game.
Send a thesis or a risk brief in plain language. Agentic pipelines evolve it into a backtested strategy through generations of graded candidates rather than a single-shot answer.
Every candidate is graded by a deterministic pipeline on probability of backtest overfitting, deflated Sharpe, and out-of-sample and walk-forward results. The proposing model never grades its own work.
The workflow
A strategy thesis or just a risk appetite, in plain language. No factor library or feature engineering on your side.
The AI proposes generations of candidates. Each one is graded on data it never saw, and the strongest survive into the next round.
A backtested strategy with its statistics and the probability it is overfit. Your user decides what to do with it.
How it grades
The proposing model never scores its own work. Every candidate is graded by a deterministic pipeline using the same overfitting-aware statistics quant funds use, and the strongest survive into the next generation. The approach is inspired by DeepMind’s research on evolving mathematical proofs.
Invents and mutates candidate strategies.
Deterministic, overfitting-aware scoring on unseen data.
The strongest candidates survive and are refined again.
The verdict goes back to your user. They decide.
The response
The edge survives out-of-sample and through walk-forward. This one earns its place.
Sharpe deflates from 2.4 to 0.2 once we account for how many variants were tried. Don’t risk real money on it.
Both answers ship in the same structured format, so your agent can relay either. Figures are illustrative.
The interface
Lagias will speak Model Context Protocol over streamable HTTP, so any MCP client can connect with a few lines of config. The shapes below preview what we’re building. Names will change before launch.
your client’s mcp config{
"mcpServers": {
"lagias": { "url": "https://mcp.lagias.com" }
}
}create_research_runStarts an evolutionary research run from a plain-language brief.
get_run_statusReports the current generation and the strongest candidate so far.
get_verdictReturns the graded strategy with its statistics and the probability it is overfit.
list_runsEvery run your user has commissioned, with verdicts.
{
"brief": "Momentum on US tech. I can
stomach drawdowns.",
"universe": "US large caps"
}get_verdict · result{
"strategy": "9-factor momentum blend",
"verdict": "LIKELY_OVERFIT",
"prob_overfit": 0.81,
"in_sample_sharpe": 2.4,
"deflated_sharpe": 0.2,
"out_of_sample": "edge fades",
"summary": "Sharpe deflates from 2.4
to 0.2 once we account for how many
variants were tried."
}A draft of the interface, not a spec. Figures are illustrative.
Aligned incentives
We’ll charge a flat subscription, never a cut of trades or assets. A broker earns when your user trades and a robo earns on their assets. Our verdict earns nothing by flattering a strategy.
FAQ
Lagias is a quant research harness for retail traders and their AI agents. It evolves plain-language strategy ideas into backtested strategies and grades every candidate with a deterministic evaluation pipeline, so the verdict tells you how much to trust the result.
It will expose the same research runs the product uses. Your agent sends a brief, the harness evolves and grades candidates, and your agent gets back the strategy with its verdict. It will connect to ChatGPT, Claude, broker agents, and custom stacks.
Language models are unreliable judges of their own output. 62% of US retail investors already use AI to inform investment decisions, yet only 23% mostly or completely trust its output (Investing.com survey, April 2026). A deterministic pipeline the model can’t game is how we close that gap.
Four checks run on every candidate. Deflated performance, probability of backtest overfitting, out-of-sample testing, and walk-forward validation. Grading always happens on data the proposing model never saw.
No. Lagias is a research and education tool, not a broker. It will not place orders and it does not give personalized investment advice. Your user makes every decision.
The waitlist is open and alpha testers are invited on a rolling basis starting August 2026. The MCP server and a hosted chat launch in September 2026.
Early access
The waitlist is open. Alpha testers are invited on a rolling basis starting August 2026; the MCP server and a hosted chat launch in September.
Backtested and hypothetical results are not indicative of future performance. For research and education, not investment advice.