Guided demo
Each scenario creates a fresh simulated user with $100,000, sets its safety policy, and plays the steps on the real pages. Pause, step or take over at any point.
Basics
Protections most brokers already offer in some form.
Fat finger
400 shares instead of 40. The review screen warns, the guard holds it, the owner rejects it.
Elsewhere: order size limits (e.g. Interactive Brokers' precautionary settings). New here: hold for the owner instead of refusing.
Limit through the market
A buy limit 10% above the ask is refused, with a one-click sensible limit.
Elsewhere: price caps and exchange price bands.
Stop on the wrong side
A buy stop below the market would fire instantly. Refused.
Elsewhere: brokers often reject stops that would trigger at once.
Stop-market made safe
A protective stop-market becomes a stop-limit with a 3% floor.
Elsewhere: some retail brokers collar market and stop orders.
The same order twice
A resent identical order is refused; the first one stays.
Elsewhere: duplicate-order warnings and pre-trade checks.
Loosening a limit
Over the owner's cap: refused. Raising the cap needs a second, explicit step.
Elsewhere: account limits are standard. Here only the owner can change them.
Agentic
An AI agent in the Agent chat makes a typical agent mistake, or the owner's guardian agent catches one. The agents are scripted; their API calls and the broker's answers are real.
Runaway agent
A looping agent resends orders. Order 6 in 5 s freezes the account; the agent stops.
Agents can loop at machine speed. A freeze plus a human beats a throttle.
Dollars as shares
Told $1,000, the agent sends 1,000 shares. Refused; it converts and buys 6.
Unit mix-ups are typical for LLM tool calls.
Retry after a timeout
A lost reply makes the agent resend. The copy is refused.
Timeouts plus retries double positions.
Stale position
The agent sells 100 it thinks it holds; the account has 50. The guard sells 50.
Agents act on memory; the guard checks it against reality.
An extra zero
40 becomes 400. Held; the owner approves 40; the agent learns the outcome.
A human in the loop without blocking the intent.
A large order, worked
The agent previews 1,200 shares, sees the plan beat sending it all now, and lets the engine slice and route it.
Elsewhere: brokers' algos (TWAP, VWAP, smart routing). New here: the agent sees the predicted cost and the engine's reasons before it sends anything.
Guardian: a drifting agent
The agent declared long ACME, then went short. The thesis score rises; the guardian proposes the LLM-agent pack; the owner applies it.
An owner can't read every order of every agent. A guardian agent watches the models and proposes; the owner decides.
Without JavaScript, Play still sets the scenario up and signs you in; you then follow the steps yourself.