What Mattered While Building Mercury Command

nahtnam

@nahtnam

·Follow

Mercury Command launched today. I led the engineering, and it changed how I think about agent products. From the outside, it's a chat box that can move money (with your explicit approval). The chat box is the smallest part of the system. Here are the decisions that mattered:

Mercury

@mercury

Mercury Command is live. AI built directly into your Mercury account that completes financial work. Natural language, real actions, and you approve everything. This is how banking* works now.

Watch on X

nahtnam

@nahtnam

·Follow

Authority. The model proposes. The product enforces. The user authorizes. The AI can prepare a payment, but it can't send one. You review the real details and confirm, and it goes through the same approvals and permissions as anything else in your account.

nahtnam

@nahtnam

·Follow

Chat is the wrong abstraction. Fine for questions. For money, you need to see the real thing rendered from live data, not the AI's description of it. Text can read as confident and still be wrong. The chat is the entry point, not the whole product.

nahtnam

@nahtnam

·Follow

No stale copies. The AI doesn't keep your account data, it points at it ("this invoice," "that account"). The product looks up the live version the moment it acts, because balances and permissions change between reading something and doing it.

nahtnam

@nahtnam

·Follow

One giant prompt didn't work. We tried cramming every product's rules (payments, cards, invoicing, treasury) into one system prompt. The instructions competed and the model got less reliable. Now it loads just the playbook for what you're asking, nothing else.

nahtnam

@nahtnam

·Follow

Old chats never die. We save your conversations, and old ones still have to work. A chat from months ago, full of old AI actions, has to load after we've changed the code behind it. Every saved chat is secretly an API contract with your future self.

nahtnam

@nahtnam

·Follow

The lasting part of Command isn't the model. It's everything around it: permissions, staged actions, real product UI, live data, and saved history. Let the model handle language. Let the product handle anything that touches your money.

nahtnam

@nahtnam

·Follow

I'll be sharing more this week, going deeper on each of these and some of the other technical decisions behind Command. Follow along if this is your kind of thing.