An autonomous agent trading the market, in the open.

Instruments

Everything the agent traded or watched in this window, named, on one shared axis. Run the pointer across and every reading moves to the same day.

Inside the agent

A quiet curve can mean an idle system or a disciplined one. The difference is here: what it looked at, what it refused, and what it currently believes.

Specification

Scope

A trading agent that runs on US equities during real market hours, on simulated capital. It is an experiment, not a product: nothing here is for sale and there is nothing to sign up for.

Autonomy

Nobody approves its trades. It reads the market, forms a view, decides how much to risk and places the order on its own. A human sets the limits it is not allowed to cross, and then stays out of the way.

Decision

Several AI models work on the same question with different jobs — one analyses, one argues against, one summarises. None of them places an order: execution is ordinary deterministic code, which is what makes it auditable.

Context

News, company events and macro indicators enter the decision as classified context rather than as free text, so the same event is read the same way twice.

Validation

No rule reaches live capital without surviving out-of-sample backtests and overfitting checks. Most candidate rules die here.

Risk

Exposure limits, automatic circuit breakers and broker-side protections. The system is built to stop itself, which it has done.

Audit

Context, decision, plan, order, execution and outcome stay linked to one another. Any single trade can be reconstructed after the fact.

Review

Every day the agent grades what worked and what did not, and adjusts its own weights. Every change is logged.

Disclosure

The instruments are named here, open or closed. Position sizes and the selection method are not published — what is under test is how the system behaves, not a list of tips.