Inside MarketPlay's AI control room: 10 agents, one approval gate

MarketPlay is a financial literacy game, but the backend that runs it is closer to an AI operations team. Ten scoped agents handle the work: generating scenario copy, grading checkpoint tests, monitoring the paper-trading broker, flagging anomalies across the six-tab analytics dashboard, and more. None of them ship a change or take an action without passing through a human approval gate first.

I built it this way on purpose. The product deals with a teenager's financial decisions, even in a paper-trading environment, so the cost of an unreviewed AI mistake is higher than in a typical side project. The control room is the compromise: agents do the grunt work at machine speed, I stay the single point of accountability for anything that touches a user.

Each agent is scoped narrowly rather than given broad autonomy. One agent's job is to draft the next scenario in the reel curriculum, nothing else. Another watches the paper-trading broker's order flow for anything that looks like a bug in fills or settlement. Narrow scope makes each agent easier to audit, and easier to debug when something goes wrong, which it occasionally does.

The pattern generalizes past MarketPlay. GetAITrade runs on the same philosophy applied to real trading infrastructure: broker connectivity and Telegram routing handled by AI, but every live order still passes a human verification gate before it executes. Human-in-the-loop isn't a compliance checkbox for me, it's the actual design principle that makes me comfortable putting AI in front of anyone's money.

I designed, built and shipped every layer of MarketPlay's control room solo. If you're building anything where AI touches financial decisions, the lesson I'd pass on is simple: decide what the agent is never allowed to do before you decide what it's allowed to do.