agent-trading-guardrails

The brakes for autonomous trading agents. A read-only, dependency-free pre-trade safety gate + audit ledger + kill switch you drop in front of any bot or LLM agent that can place real orders.

Everyone selling into the "AI agent that trades" space sells alpha — strategies that decay. This sells the opposite: the thing that stops your agent from blowing up the account. A safety layer's credibility compounds with every incident it prevents, instead of decaying like an edge. It is the picks-and-shovels play on the whole autonomous-trading gold rush.

It decides; it never trades. check() is pure and side-effect-free — it returns a verdict and a risk-sized dollar stake. Your code still places the order.

This engine is extracted from a real-money trading system's pre-trade rule checker — the same gate that runs in front of live prediction-market and brokerage execution.

Why it exists

An LLM wired to Robinhood / Kalshi / Polymarket / Alpaca will, eventually:

• fat-finger an order size,
• revenge-trade straight into a drawdown,
• chase a longshot at 3¢,
• or silently die mid-session with an open position and no stop.

These are not strategy problems; they are safety problems, and they are the same across every venue. This package encodes the boring, battle-tested controls that catch them — as configuration, not code you have to write.

Install

pip install -e .          # from this directory
python -m guardrails.selftest

Stdlib-only. No numpy, no pandas, nothing to audit or pin.

Quick start

from guardrails import Guardrails, GuardrailConfig, AuditLedger, KillSwitch

gate   = Guardrails(GuardrailConfig.conservative())
ledger = AuditLedger("audit/trades.jsonl")
halt   = KillSwitch("audit/HALT")

verdict = gate.check({
    "side": "YES",          # YES = "this resolves true / goes up"
    "market_price": 0.52,   # the ask, in (0, 1)
    "model_prob": 0.66,     # your model's probability of YES
    "balance": 1000.0,
    "peak_balance": 1200.0, # enables the drawdown breaker
})

ledger.record_decision(verdict, venue="kalshi", symbol="KXFED-26JUL", side="YES")

if not halt.is_engaged() and verdict.passed:
    place_order(dollars=verdict.size)      # <- your code
else:
    print(verdict.blocking_rules, verdict.reasons)

The rules

Every threshold is a field on GuardrailConfig — set your own without touching engine code. Stable rule codes (R-0xx) are what your ledger and dashboards key on.

edge_pt = (win_prob − market_price) × 100, where win_prob is model_prob for a YES buy and 1 − model_prob for a NO buy.

Trade dict — keys

Required: side ("YES"/"NO"), market_price (0–1), model_prob (0–1), balance ($).

Optional (each enables a rule; absent ⇒ that rule is skipped, logged as an advisory): edge_pt, open_position_cost, bankroll, market_volume, peak_balance, recent_daily_pnl (most-recent-first), today_pnl_dollars, starting_balance_dollars, strategy. Unknown keys are ignored, so you can pass your own trade objects through untouched.

Works on any venue

• Prediction markets (Kalshi, Polymarket) — native; market_price is the contract price. See examples/kalshi_example.py.
• Brokerage (Robinhood, Alpaca) — map a directional thesis onto a win probability and a break-even "price" from your reward-to-risk; verdict.size is dollars-to-risk, convert to shares with your stop distance. Disable the R-014 anti-longshot floor for this mapping (it's a prediction-market rule) — see examples/robinhood_example.py.

The audit ledger

AuditLedger writes an append-only JSONL row for every decision and every fill — the compliance artifact that makes an autonomous strategy defensible. Writes never raise: a logging failure returns False instead of blocking your order path.

The kill switch

KillSwitch is a file-based, out-of-band halt. Any human or monitor can create the halt file and the next is_engaged() refuses new entries — no process, no network, no coordination with the agent. Keep your exit path unconditional: halt entries, never exits.

Testing

python -m guardrails.selftest     # zero-dependency smoke path
python -m pytest tests/           # full unittest suite

License

MIT. Not financial advice — a risk-control tool, not a guarantee against loss.