AI Agent · Trading · 2026
Sentinel-X
The market runs on emotion. This doesn't.
An autonomous momentum trader on Robinhood's agentic MCP - deterministic math picks the trades, Claude is only allowed to say no, and exchange-side stops guard every position while the agent sleeps.
2
runs per trading day
26
validator tests before any action
60
shadow days before $1 at risk
The problem
Retail trading fails for human reasons: emotion at the entry, hope at the exit, inconsistency everywhere in between. And most "AI trading" makes it worse - it hands a persuasive language model the keys and calls conviction a strategy. The interesting question was never whether an agent can trade smarter than a person. It's whether an agent can trade more honestly: same rules every run, every decision written down, no benefit of the doubt.
The approach
Sentinel-X is built on a separation of powers. A deterministic Python core decides what is tradeable - EMA crossovers, 20-day relative strength against SPY, volume confirmation, position sizing, stop levels - and its 26-test suite must pass before any run is allowed to act. Claude sits above it with exactly one power: veto. It reads the news behind each candidate and can reject a trade whose momentum is really an acquisition bid or a binary event, but it cannot add a symbol, change a size, or move a stop. Every live entry ships with an exchange-side stop that Robinhood enforces while the agent is offline, and account-level circuit breakers - daily loss, losing streak, weekly drawdown - can halt the whole system. So can one email: reply HALT to any digest.
Architecture & build
Two scheduled Claude runs fire each trading day (10:00 and 15:30 ET) inside a ring-fenced Robinhood agentic account - the broker's API firewall makes the primary portfolio untouchable at the protocol level. All state lives in versioned, human-readable documents: an append-only trade ledger, a position book, a config that is law, an incident log. After each close, results publish through a GitHub relay to a self-hosted dashboard behind Cloudflare Zero Trust, and a daily digest lands by email and push. The schedule even adjusts itself for daylight-saving transitions.
The outcome
Live in shadow mode since July 2026: the full pipeline runs on real market data and writes hypothetical fills to the ledger, but no real order exists until the system passes a statistical gate - thirty completed round trips, returns at or above SPY net of modeled slippage, drawdown under ten percent, and zero integrity failures across sixty trading days. The ledger is the product; profits are a hypothesis under test.
- Broker
- Robinhood Agentic MCP (ring-fenced account)
- Cadence
- 10:00 + 15:30 ET, cloud-scheduled
- Risk
- 5% max per position, -8% exchange-side stops, 3-layer breakers
- Gate
- 60 shadow days, beat SPY, zero integrity failures
Personal research systems, not investment products. Nothing here is financial advice.