When should a task you run through your agent become an app?

TL;DR

When the same work goes through your agent every week, have the agent build it as an application instead. The app runs on a schedule, gives the same answer every time, costs a fraction of the tokens, and the team opens a URL instead of asking you. floo gives that application a URL, sign-in, data, and a schedule, so it keeps running without you.

Three signs it is time: the request repeats on a schedule, other people need the result, or the answer must not drift.
The cost gap is real. A financial desk replaced daily Claude jobs costing about $10 a day in tokens with software that runs for under $1 a day.
The agent does not go away. It builds the app, changes it, and calls it with its own key.
The app needs four things to run without you: a URL, sign-in, data, and a schedule.

Asking your agent is the fastest way to do something once. It is a slow and expensive way to do the same thing every morning. The second time you paste the same prompt, the work has started to become a program.

How do you know a task should be an app?

Look for any of these:

It repeats on a schedule. Every morning's numbers, every Friday's report, every hour's check of who needs attention.
Other people need the result. If the output lives in your chat history, everyone else has to ask you for it.
The answer has to be the same twice. A model asked to rebuild the same script every day will rebuild it slightly differently. A finance number, a stock count, or an on-call rotation cannot drift.
It touches data that should be kept. If you paste last week's result back in so the agent can compare, the data wants a database.

What did that look like for a real team?

A financial firm used AI every morning to rebuild and debug a script that pulled its numbers. The team worked from Excel with point-in-time data.

The agent built that work as software instead. It now collects the data every minute, and the whole desk signs into a live dashboard. The daily Claude jobs had cost about $10 a day in tokens. The software runs for under $1 a day. The full use case is here.

The point is not that the model got cheaper. The model stopped doing the repeated part. It did the part it is good at, writing and fixing the program, and the program did the rest.

What does the agent do after the app exists?

More, not less. It changes the app when the work changes, adds the next view someone asks for, and fixes the job that failed overnight. It can also be a user of the app: give it a scoped key on the routes it needs, and it can update records or pull results through the API while people use the UI. How to give your agent its own key.

That is the pattern we build floo for: a person defines what the app should do, an agent builds it, and floo runs it.

What does the app need to keep running without you?

Four things a script on a laptop does not have:

A URL the team can open.
Sign-in, so only the right people get in. On floo, access_mode = "accounts" makes the app invite-only, and your code reads the user from a header.
Data that persists. [managed.default] with type = "postgres" gives the app a database and injects DATABASE_URL.
A schedule. A [cron.<name>] section runs a command in the app's container, every five minutes or every Monday.

Our own CRM started this way. An hourly job pulls fresh product data and sends the next actions to Slack, and the team signs in to see the rest. How we built it.

When should it stay a conversation?

When the task is one-off, when it needs judgment each time, or when the inputs are different every run. Drafting a reply to a customer, investigating an incident, or exploring a new data set are agent work. Turning the same export into the same chart every morning is not.

A useful test: if you could write down the steps and they would be the same next week, the agent should write them down as code.

Sources

floo use cases: financial dashboard, floo-crm
floo docs: app sign-in, Postgres, cron jobs, app API keys

Frequently asked questions

When should I stop asking my agent and build an app instead?

When the same request comes back on a schedule, when other people need the result, or when the answer has to be the same every time. Those are signs the work is a program, not a conversation.

Is running software cheaper than running the same task through a model?

Usually, for repeated work. One financial desk replaced daily Claude jobs that cost about $10 a day in tokens with a dashboard on floo that runs for under $1 a day.

Does the agent still have a role once the app exists?

Yes. The agent builds the app, changes it when the work changes, and can call it through a scoped API key. The app does the repeated part deterministically.

What does the app need to keep running without me?

A URL, sign-in for the people who use it, somewhere to keep data, and a schedule. On floo those are declared in floo.app.toml: access_mode, a managed Postgres database, and [cron.<name>] jobs.