Lab experiment · Building in public

We gave an AI one trade a day.

An autonomous agent that picks a single momentum trade each market day, with real money on the line and every dollar of risk controlled by deterministic code it cannot override. The AI decides what. The math decides how much, and when to get out. This page shows it running, live.

Live status

What the agent is doing right now.

State
Loading…
Position
Trail
Waiting for the first published status…
Updates every few minutes during US market hours (9:30–16:00 ET). Dollar amounts are never published; this is a percent-terms view of a small experimental account.
How it works

The AI only answers one question.

Most trading bots fail by doing too much. This one splits the job three ways and lets each part do the thing it is actually good at.

01

The AI picks

Once a day, after the open settles, the agent reviews which names in a fixed ten-stock universe are actually moving, reads the day's catalysts, and returns one pick with a conviction rating. Or it passes. Passing is a feature, not a failure.

02

The math decides size and exit

Position sizing, the protective stop, the ratcheting trailing stop, the daily loss breaker: all plain deterministic Python. The AI cannot touch any of it. Every worst case is knowable before entry.

03

A second brain must agree

Before any order, the pick is checked against Signa, an independent quant engine where 30+ models vote nightly on every name. Live signal at 70%+ confidence, nightly consensus at grade B or better, or no trade today.

The stack

Four pieces, no cloud.

The whole thing runs on a Mac mini in a closet. That is the point: agentic infrastructure is now small enough to sit next to your router.

Claude

Makes the daily pick and executes broker calls through the Claude Code CLI. Reads the news before it decides.

Robinhood Agentic API

The official MCP rail for agents, launched May 2026. Runs against a dedicated cash account that is walled off from everything else.

Signa

Independent 30-model quant consensus with letter grades. The bot's veto layer. Two systems must agree or nothing happens.

Deterministic Python

One auditable file owns every dollar decision: sizing, stops, breakers, reconciliation. Forked from an MIT open-source project, audited line by line.

Guardrails

Rules the agent cannot break.

Why we built it

The interesting question was never "can an AI trade." It was "can you wrap an AI in enough deterministic structure that you would let it act unsupervised." That question applies to every agent we ship.

Innovative Group lab notes
Honest fine print

This is an experiment.

It trades a deliberately small account funded with money we can fully afford to lose. It holds overnight, and overnight gaps can blow through any stop. The AI can be wrong. The quant consensus can be wrong. Both can be wrong together.

Nothing on this page is financial advice, a track record, or a product you can buy. We publish it because we think building agents in the open beats writing thought-leadership about them. The bot is a fork of an MIT-licensed open-source project by JC Merlo, extended with the Signa gate and our own guardrails.

The same idea, applied to your business

Agents with guardrails are the product.

Next Best Action is what this architecture looks like pointed at an agency-client relationship: AI does the reading, deterministic structure does the deciding, and a named human owns every consequential action.