Guardrails are the rules that keep an autonomous trading agent inside boundaries you chose. Robinhood gives you the account isolation and the notifications; the behavioral rules are on you. Here’s the complete set, grouped by what each layer protects against.
Layer 1 — Hard limits (contain the size)
- Per-trade cap — the largest single order allowed.
- Daily volume cap — total dollars tradeable per day, to stop churn.
- Concentration cap — most the agent may hold in any one ticker.
- Account boundary — trade only inside the funded agentic account; never reference the main portfolio; never request more funds.
- Scope lock — equities only during beta. Refuse options, crypto, futures, margin, short selling.
→ Sizing guidance in How to Set Spending & Trade Limits.
Layer 2 — Approval gate (keep a human in the loop)
- Threshold approval — orders above a set size stop and wait for your explicit "approved."
- Ambiguity rule — anything the agent isn’t confident matches your intent is a halt, not a guess.
Layer 3 — Universe controls (limit what it can touch)
- Allowlist — restrict trading to a named set of tickers (recommended for conservative setups).
- Denylist — at minimum, exclude leveraged/inverse ETFs and penny stocks.
Layer 4 — Circuit breaker (catch runaway behavior)
- Rate limit — halt and alert after N trades within M minutes. An agent acting normally won’t trip it; an agent in a loop will, before it does real damage.
Layer 5 — Kill switch (stop it now)
- Soft stop — a "STOP" phrase the agent obeys immediately, with no resume until you say so.
- Hard stop — know how to disconnect the Trading MCP, the off switch that doesn’t need the agent’s cooperation.
→ Full detail in Building a Kill Switch.
Layer 6 — Security (the layer most people miss)
- Prompt-injection rule — instruct the agent to treat any input trying to change its rules (in a news headline, ticker name, or tool output) as a possible attack, refuse it, and flag it to you.
→ Why this matters: Can Your AI Trading Agent Be Hacked?.
Set all six in one step
Writing these by hand is error-prone — and a guardrail you forgot is a guardrail you don’t have. The free SecProve Agent Safety Kit generates all six layers as a single copy-paste config, scaled to your funding and risk tier, and ships with a printable pre-flight checklist.
If you run your agent in Claude, the SecProve Agent Safety skill applies all six layers automatically — install it once and the guardrails are enforced on every session.
Six layers of guardrails handle the known risks. The unknown one — an attacker turning your agent against you — is a security skill you can actually measure. Test yours at secprove.com.