Use cases

What people hand over to an agent

A use case here is not a prompt. It is one objective your agent owns end to end: what to read, which tools it may use, what has to be true when it is done, and who takes over when the decision is not its to make.

Start free See how work runs

Four jobs below, each written the way it runs rather than the way it demos — with the tools it uses and the point where a person takes over.

You hand over: An objectiveOne outcome with a done condition — not a chat you have to supervise.
It brings: ToolsThe tools you allow, used within your permissions and budget.
You get back: A recordWhat changed, who allowed it, and whether it stuck.
You keep: The exceptionsDecisions you have not delegated arrive as a handoff with the context.

What a use case means here

An objective your agent owns end to end.

Not a prompt you paste and supervise. Every job below answers the same four questions — and where one has no answer, that is the first thing the work tells you.

See how work runs

The situation

One line about the moment the work starts: a message arrives, two systems disagree, a check nobody has time for comes around again. If the situation needs a paragraph, it is two jobs.

What it does

The sources it reads, the tools it uses, and the rule it decides with. Writing these down is what makes the work checkable later — that is the difference between a prompt and a job.

What you see afterwards

The result in your terms: a message the right person actually received, a booking confirmed by the site that took it, a calendar that now matches the plan. You should be able to point at it without asking what was done.

Example · family schedule

job_7C21

objective: move piano to Thursday
actions: schedule updated · 2 messages sent
decision: Grandma free Thursday · asked, agreed
result: Thu 17:00 · pickup by Grandma
confirmed by: family calendar · Telegram

Illustration. Example ids, values, and amounts — not data from your workspace.

Where a person steps in

The boundary, written before the first message: the threshold, the exception, the decision that stays yours. Handing over is a normal outcome, not an error — when you take over, the agent stops and passes on everything it knew.

Four jobs

The work people hand over first.

One job each for personal life, family, school, and work. Each row is one objective: what the agent reads and does, what you see afterwards, and where a person takes over.

Job 01

Personal secretary

A flight moves by a day, and three plans hang off it.

Web · Memory · Messaging · Scheduling

What it does: Reads the airline email you forward, finds what depends on the old date in your calendar and your notes, and proposes the changes. Moves what you allowed, drafts the messages to the people affected, and sets reminders for the new plan.
What you see afterwards: An updated calendar, the messages that went out, and reminders scheduled for the new dates, each with a record you can read back.
Where a person steps in: Anything that costs money or speaks for you to someone new: the agent prepares it, you approve it on a card.

Tools it uses

Read web pages
Search memory
Send messages
Schedule reminders

Job 02

Family assistant

A lesson moves, and the pickup plan for the week has to change with it.

Messaging · Memory · Scheduling

What it does: Reads the message in the family group, updates the shared schedule, and works out who now needs to be where. Tells the people affected in the same chat and remembers the new routine, so next week starts from it.
What you see afterwards: The changed schedule, the message each person received, and a memory entry for the new routine that anyone in the family can correct.
Where a person steps in: Payments, anything about a child’s health or safety, and any plan two adults disagree on: it asks the people who decide.

Tools it uses

Read messages
Send messages
Save to memory
Schedule reminders

Job 03

Teaching assistant

Friday’s quiz needs writing from this week’s notes, and the class needs reminding.

Files · Memory · Messaging · Scheduling

What it does: Reads the notes and slides you point it at, writes the quiz and an answer key at the level you set, and schedules the reminder to the class channel for the time you choose.
What you see afterwards: The quiz and answer key as files, the reminder as a scheduled message with a delivery record, and the notes each question came from.
Where a person steps in: Grades, anything said to a student about their progress, and every message to the class before it is sent: the teacher approves.

Tools it uses

Read files
Write files
Send messages
Schedule messages

Job 04

Customer support

A delivery is late and the customer wants the express fee back.

Messaging · Memory

What it does: Reads the conversation, looks up the order and your refund policy, and decides whether the fee is refundable. Replies to the customer on the same channel and remembers the decision, so the next similar question starts from it.
What you see afterwards: A reply the customer received, the refund confirmed with the payment provider, and a memory entry noting the policy rule that decided it.
Where a person steps in: Anything above your dispute threshold, anything the policy does not cover, and anyone who asks for a person: a handoff with the whole conversation attached.

Tools it uses

Read messages
Search memory
Send replies
Save to memory

These four are patterns, not case studies: no customer, no logo, no quote, and no number invented for effect.

Tool groups you can turn on per agent: Computer — run commands and edit files in an isolated workspace. Web — search and read public web pages. Memory — remember and look up facts. Messaging — send, react to, and read messages. Scheduling — follow-ups and recurring work.

Choosing the first job

Repeatable, reversible, measurable.

The first job is not the most impressive one. It is the one where a wrong answer costs an apology instead of a lawsuit: start where being wrong is cheap to correct.

Repeatable

The same situation arrives on its own, and you can write the steps down in order. If every case needs a different answer, you have judgement to encode before you have a job to hand over.

Reversible

The worst outcome can be undone: a refund, a correction, a deleted message, a rebuilt file. Start where a wrong answer is cheap to correct, and let your agent earn the irreversible work.

Measurable

You can say what a correct outcome looks like, and check it in the product rather than by feel: a reply that was received, a change confirmed by the system it was made in, a record that now agrees.

One boundary

One objective per agent, one channel to start, one person to escalate to. Two boundaries make a job look finished while half of it quietly stays with you.

The one-sentence test

Write the job down before you build it. If both halves of the sentence are specific, it is ready to hand over.

When <trigger> arrives, read <sources>, decide with <rule>, then do <action>. If <exception>, hand it to <person> with the context.

The same sentence, filled in

When a lesson moves, read the family chat and the weekly schedule, decide who now does each pickup, then tell them. If nobody is free, hand it to a parent with the plan so far.

Most sentences that fail this test fail in the same place: a rule that was never written down. That is not a reason to stop — it is the first thing to write.

Start here

A question you already answer the same way every time
A check whose source system your agent can read
A handoff that already has a person at the other end

Wait

A decision two people would answer differently
Spend you have not capped, or a commitment you cannot withdraw
An action you cannot verify afterwards, or cannot undo

Limits

What does not fit yet.

Three kinds of work that need a different shape before they become a job you can hand over. A boundary you cannot name is one you cannot enforce, so these are named.

Not yet · rule not written

Judgement you have not codified

If two people would answer the same message differently, the rule does not exist yet. Your agent can draft the answer, but it cannot own a decision you have not defined. Write the rule down first, or keep a person on every decision.

Not yet · physical world

Actions that reach into the physical world

Every tool on this page is digital: computer, web, memory, messaging, scheduling. Nothing drives a vehicle, opens a door, or picks something up. An errand in the physical world is a handoff to a person.

Not yet · no reversal

Actions you cannot undo

Every important action is checked, confirmed, and reversed where a reversal exists. Where nothing can be undone — a payment that cannot be recalled, a message that cannot be unsent — keep a person in charge: the agent prepares, a person approves. A permission that was denied stays denied.

None of this is a dead end. Each of the three starts as something you can change: a rule you write down, a tool you allow, or an action that stays with a person. Until it changes, your agent can still gather the context and draft the work — the decision just stays yours.

Hand over one job, not all of them.

Pick the repeatable, reversible one and get it working: one agent, one channel, one real reply you can read back — and a record of what it did.

Start free Read the quickstart

Illustrations are examples, not data from a real workspace.