Human Approval for AI Trading Agents

An AI trading agent that can size and place orders on its own is one bad signal away from a real loss — this shows how to gate order execution behind a human yes.

The problem is the order, not the analysis

Letting an agent scan news, compute indicators, or backtest a strategy is low-risk — nothing happens until an order goes to a broker. The risk shows up exactly at that boundary: a mispriced signal, a stale data feed, or a hallucinated position size turns into a real fill in seconds, and unlike a bad blog post, a bad trade cannot be quietly deleted afterward.

The fix isn't slowing the agent's reasoning down — it's putting a stop between "the agent decided" and "the broker executed." That stop needs to be a real dependency the code cannot skip, not a comment telling the model to "confirm with the user first."

Gate the broker call, not the strategy loop

The pattern is the same three calls as any other Impri integration: push the proposed order as an action, wait for a decision, then call the broker's API only if approved. The strategy code that decides what to trade never talks to the broker directly — a thin wrapper does, and that wrapper refuses to run without an approved status.

python

import os
import time
import requests

IMPRI_KEY = os.environ["IMPRI_API_KEY"]
BASE = "https://api.impri.dev"
HEADERS = {"Authorization": f"Bearer {IMPRI_KEY}"}

def propose_order(symbol, side, qty, est_price):
    body = {
        "kind": "trade.order",
        "title": f"{side.upper()} {qty} {symbol} @ ~${est_price}",
        "preview": {
            "format": "markdown",
            "body": f"**{side.upper()} {qty} shares of {symbol}**\n\n"
                    f"Estimated price: ${est_price}\n"
                    f"Estimated notional: ${qty * est_price:.2f}\n\n"
                    f"Signal: 20/50 EMA crossover, RSI 34",
        },
        "idempotent": False,
        "undo": f"Submit an offsetting {'sell' if side == 'buy' else 'buy'} "
                f"order for {qty} {symbol} to flatten the position",
        "expires_in": 300,  # a stale signal is a wrong signal, keep this short
        "editable": ["preview.body"],
    }
    r = requests.post(f"{BASE}/v1/actions", headers=HEADERS, json=body)
    return r.json()["id"]

def await_decision(action_id, poll_every=5, timeout=300):
    deadline = time.time() + timeout
    while time.time() < deadline:
        r = requests.get(f"{BASE}/v1/actions/{action_id}", headers=HEADERS)
        data = r.json()
        if data["status"] != "pending":
            return data
        time.sleep(poll_every)
    return {"status": "expired"}

def place_order_if_approved(symbol, side, qty, est_price, broker_client):
    action_id = propose_order(symbol, side, qty, est_price)
    decision = await_decision(action_id)

    if decision["status"] != "approved":
        print(f"Order not sent: {decision['status']}")
        return

    # broker_client is your own SDK/wrapper — never called before this line
    order = broker_client.submit_order(symbol=symbol, side=side, qty=qty)

    requests.post(
        f"{BASE}/v1/actions/{action_id}/result",
        headers=HEADERS,
        json={"status": "executed", "payload": {"broker_order_id": order.id}},
    )

Note the short expires_in: 300 seconds, the API minimum. A trade signal that's five minutes stale is a different trade — treat expired the same as rejected and require the strategy to re-propose with fresh prices rather than resurrecting an old approval.

What the approval card should show

Field	Why it matters here
Symbol, side, quantity, estimated price	The reviewer's actual decision surface
Estimated notional	Catches size errors before they're a keystroke away from real money
Signal / reasoning summary	Lets a human sanity-check the "why," not just the "what"
`idempotent: false` badge	Warns against retrying a timed-out request, which would double the position
`undo` text	Tells the reviewer the position isn't permanent if they're wrong

Keep preview.body short and numeric. A reviewer approving from their phone between meetings needs the notional value and the side, not a paragraph of technical-indicator prose.

What Impri does not do here

Impri does not evaluate the trade, check it against a risk model, or know what a "reasonable" position size is for your account — it stores the proposal, notifies you, and holds the decision until you act on it. Your strategy code still owns risk limits, position sizing, and market-data validity; Impri only owns the question of whether a human said yes before the broker call fired. And it's a real gate only if the strategy process holds no broker credential capable of placing orders outside this path — see wrapping a tool with a human approval step for how to structure that boundary. For the general shape of the fintech vertical beyond trading specifically (payments, transfers, refunds), see human approval for fintech AI agents.

Next step: quickstart to get an API key, then adapt the wrapper above with the Python SDK instead of raw requests calls.