Three-week live test of an agent trading with real money

wakerone · reddit · 2026-07-20

A live three-week experiment running an agent on real money shows what guardrails actually matter for autonomous systems.

The author built Orus, an MCP server for Farao, and ran it on about $40 of self-custodial Hyperliquid funds using Claude Code.
The most important safeguard was removing withdrawal/transfer permissions entirely from the API surface, so even a fully compromised session can only trade within signed limits.
Trades use two-phase execution: prepare stores the order with a 5-minute TTL, then confirm executes only after either a human approval or server-side revalidation of limits.
Limits such as max leverage, per-trade margin, daily caps, and blocked assets are signed on-device and checked again server-side; the model never sees keys, only an OAuth session token revocable from the phone.
The author also used graduated autonomy: one week fully manual before allowing autonomous execution.
Exchange-native TP/SL orders were crucial; one overnight gap on a Korean index perp was stopped out for only -$1.44.

What the guardrails did not solve was fee bleed: the agent went 50W/48L in direction, which was roughly break-even on PnL but still lost more to fees than direction. One example was 7 round trips in 17 minutes for 3 cents of profit and 14 cents of fees.

The next safeguards planned are trade-frequency caps, thin-market warnings, and a weekly self-written recap from the agent.

Original post →

Three-week live test of an agent trading with real money

More from coding & agent