Guide

Review before place

The human-in-the-loop gate before an agent can spend, trade, or move money. Agent proposes; you confirm; then place. Pair it with paper before live, spend caps and a dedicated agent wallet — auto-place is the failure mode.

By Agent Lane Editorial · Page dated 25 September 2026

Answer first

Before Cursor or Grok can call a live place, pay, or signature tool, require review before place: the agent shows amount, destination, asset, and venue; you say yes; only then does the order or payment run. Stack it after a dedicated agent wallet and agent spend caps. Do not grant standing skip — see never skip review before place. After you confirm a place, keep a receipt log after the session. Agent Lane publishes the checklist. We do not place trades, hold wallets, or run your agents.

What this page is

A short operator guide for the human gate on the shield — the step that sits between caps and the venue. Pair it with the brokerage shield, the crypto shield, x402 buyer basics, the trading-agent map, and the shared-box checklist when the agent runs on a shared computer.

We are a publisher, not a trading desk. Product screens change; confirm tool names on the official page before you put money in.

Why auto-place is the failure mode

• Runaway loops. A retry, “try again,” or agent-to-agent handoff can re-fire a place tool without a new human look.
• Fat-finger size. Wrong notional, extra zero, or share count the model invented — filled before you read the quote.
• Wrong venue. Off-allowlist symbol, network, URL, or brokerage path that still looks “cheap” in the dump.

What “review before place” means

• Agent proposes. Clear fields: size or notional, asset, destination or symbol, venue or network, and whether this is place, pay, or sign.
• Human confirms. You read those fields. Missing, rounded-away, or buried detail is a no.
• Then place. Only after an explicit yes does the agent call the live tool. No “just place it” or “don’t ask.”

Where it sits in the shield stack

1. Dedicated agent wallet — money the bot can reach; long-term holdings stay outside.
2. Spend caps — session and daily maxes you enforce; funded balance is the last brake.
3. Review before place — human gate on amount, destination, asset, and venue.
4. Venue — brokerage MCP, on-chain pay path, or other rail behind the gate.

Skip a layer and the next one has to catch everything. Stack all four before the first live call.

How to require review before place

1. Ban auto-place in standing rules. Write rules the agent must keep: no “just place it,” “don’t ask,” or pay-on-quote without a human confirm.
2. Require a clear proposal. Agent must show size, asset, destination or symbol, and venue or network before any place, pay, or signature tool runs.
3. Confirm, then place. You read the proposal. Only after an explicit yes does the agent call the live tool.
4. Enforce size and venue limits. Cap order size and notional. Allowlist venues and destinations. Refuse unknown URLs, symbols, or networks.
5. Keep a kill switch ready. Know how to disconnect live connectors, lock the agent apps, and empty or move the spend sleeve in one pass.
6. Revoke after the session. When the run ends, disconnect connectors, revoke grants, rotate tokens, and empty leftover funds before the next raise. Before you empty or raise: write a receipt log after the session (amount, destination, asset, venue, connector, time).

Practical checklist

• Size limits. Max notional and max size per place — in code or settings you control, not only a chat reminder. Pair with agent spend caps.
• Allowlist venues. Known brokerage paths, networks, pay-to addresses, and symbols only. Unknown destination is a no.
• Kill switch. One pass: disconnect connectors / payer MCP, lock or sign out of agent apps, empty or move the spend sleeve.
• Revoke after session. Grants and tokens that lived beside those connectors get revoked or rotated when the project ends — see revoke before you debug, dedicated agent wallet, and shared-box. Proof of what ran: receipt log after the session.

Brokerage path: Shield brokerage spend. On-chain pay: Shield crypto spend and x402 buyer basics.

What Agent Lane does not do here

• We do not place trades, preview your orders, or babysit your agents.
• We do not hold your funds, run a hot wallet, or settle payments.
• We do not publish a funded “agent buyer” or a set-it-running auto-place prompt.

Related guides

• Robinhood MCP: connect an external AI agent
• Dedicated agent wallet
• Agent spend caps
• Never skip review before place
• Session time-box
• Revoke before you debug
• Receipt log after the session
• Separate browser profile for the agent
• Fund the session ceiling
• One connector at a time
• Paper before live
• Connect only what you need
• Shield crypto spend
• Shield brokerage spend
• x402 buyer basics
• Trading-agent map
• Write invalidation before entry
• Reject duplicate place same thesis
• Confirm symbol before place
• No scale-in without GO
• Flatten all before connector revoke
• One order per idea until fill
• No place while unconfirmed fill
• Write max loss before entry
• Shared-box and connector checklist
• Agent memory hygiene
• More on Resources

FAQ

What does review before place mean?

Review before place is the human-in-the-loop gate: the agent proposes amount, destination, asset, and venue; you confirm; only then does a place, pay, or signature run. No “just place it,” “don’t ask,” or auto-place on a quote.

Why is auto-place the failure mode?

Auto-place turns a bad prompt, runaway loop, fat-finger size, or wrong venue into a live spend before you see it. Caps still let a bad order fill inside the ceiling. Separation of wallets does not stop the first wrong place. The review gate is what stops the fill until a human says yes.

Where does review sit in the shield stack?

Stack order: dedicated agent wallet (money the bot can reach) → spend caps (how much) → review before place (human gate) → venue (brokerage, on-chain pay, or other rail). Skip a layer and the next one has to catch everything.

What should the agent show before I confirm?

At minimum: size or notional, asset, destination or symbol, venue or network, and whether this is a place, pay, or signature. Refuse if any field is missing, rounded away, or buried in a long dump.

What belongs on the practical checklist?

Size limits you enforce, an allowlist of venues and destinations, a kill switch (disconnect connectors / empty the sleeve), and revoke after the session. Pair with agent spend caps and a dedicated agent wallet.

Does review replace caps or a dedicated wallet?

No. Review stops the first bad place; caps bound how much can move if a rule fails; the dedicated wallet keeps long-term holdings outside. You need all three before a live connector can spend.

Where do I apply this next?

Hard dollar limits: agent spend caps. Separate spend path: dedicated agent wallet. Proof after place: receipt log after the session. Brokerage MCP: shield brokerage spend. On-chain hygiene: shield crypto spend. Paying for a web request: x402 buyer basics. Shared computer: shared-box checklist. Map of trading agents: trading-agent map. Agent Lane publishes the checklist; we do not place trades or hold wallets.

Next steps

Before an agent touches real money, shield your brokerage account, map the trading-agent options, and set hard spend caps. Or browse all Agent Lane guides.