How we taught Macy to use (nearly) every tool on the internet

When you text Macy “is 6E 2044 on time?”, she doesn’t actually know anything about flights. What she knows is how to find something that does — while you watch the typing indicator, inside the same thread where you asked. That trick is most of the product. Here is how it works.

One agent, one socket

Early on we made a decision that has paid for itself every week since: Macy is one agent, and she talks to every tool through one interface. Flights, weather, your inbox, a meme generator — from her side they all look identical. A tool has a name, a description of what it does, and a way to call it. Nothing else about the outside world leaks in.

The tempting alternative is to hand-build a feature for every use case. A flights feature, a weather feature, a stocks feature — each one polished, each one a little island, each one another thing to maintain forever. With one interface, adding a capability doesn’t mean building a feature. It means plugging one more tool into a socket that already exists.

Your apps are different

The tools that touch your life directly — your inbox, your calendar, your docs, your team chat — don’t get treated like the rest. Each connection is authorized by you, for you, through the app’s own standard sign-in. Macy holds keys for your account and nobody else’s.

And she acts with a hand brake on. Reading is quiet: she can check your inbox or your calendar without asking. Acting is loud: anything that sends, books, posts, or spends waits for your yes, in the thread, every time. It’s the difference between an assistant reading the mail on your desk and one mailing letters in your name. The second one should ask.

The long tail is a marketplace

The interesting question was never email. It was the ten-thousandth capability — the transcript of a cooking video, the review history of a biryani place, the traffic numbers of a website you’re curious about. No team can hand-build that list. So we stopped trying.

Instead, Macy searches a live marketplace of thousands of tools at runtime. When you ask for something she has no built-in answer for, she looks up tools that claim to do it, reads their descriptions the way you’d skim search results, and picks one — with a strong bias for the cheapest tool that plausibly does the job. Not the fanciest. The cheapest that works. If it fails, she has learned something the description didn’t say, and she picks differently next time.

The first time we watched her find and use a tool nobody on the team had ever heard of, it felt less like running software and more like watching someone figure out a new kitchen.

Costs that cannot run away

Every tool call is metered. Each person gets a small daily budget — genuinely small, a number we’d be comfortable losing entirely — and every call Macy makes is charged against it. When it’s gone, she says so and stops, instead of quietly burning money.

The subtle part: failed calls cost money too. A tool that errors still bills. So Macy never blindly retries. A failure is a signal to stop and think — try a different tool, ask a clarifying question, or admit this one isn’t working — not a signal to hammer the same endpoint five more times. A surprising amount of responsible agent design falls out of one rule: waste has a price.

Slow is fine. Silent is not.

Some tools are slow. Transcribing an hour of video takes a while no matter who does it. So Macy never goes dark: if something will take more than a moment, she says so — “this one’s taking a minute, i’ll text you when it’s done” — and then actually does. A thread where the other side tells you what’s happening is the difference between texting a person and filing a ticket.

Shaped by you

The marketplace means Macy’s edges are soft. There will always be things she can’t do yet, and the honest answer to “what can she do?” changes every week. So the feedback loop lives in the same thread as everything else: text her what you wish she could do, and it lands in front of us. A surprising amount of what she does today started as someone’s annoyed message at 11pm.

That’s the system. One agent, one socket, thousands of tools, a budget that can’t run away, and a feedback loop that is just… texting. It’s the least glamorous architecture diagram we’ve ever drawn, and the most useful one.