Agents
Sign up

Turn ideasinto working agents.

Describe a job the way you'd explain it to a new colleague. Agents works out when to start, which tools it needs and where to stop and ask you.

The plan updates as you type.

Agent plan

Draft

Start typing. Your agent's plan takes shape here: what kind of agent it is, when it starts, what it uses and where it checks with you.

An agent is a job description that carries itself out.

A chatbot waits for a question and answers it. An agent has a job. It knows when to start, what it's allowed to use and when to hand the work back to you. Every agent you make here has the same six parts.

Goal
The outcome you want, in a sentence. The agent decides the steps and changes them when something unexpected turns up.
"Overdue invoices get a reminder within a day."
Trigger
What sets it off: a time, an event or you asking. Without a trigger it's a prompt you have to remember to run.
Every Monday at 9. A new email. A changed row.
Tools
The places it can read from and act in. An agent can only touch what you connect, and you choose read-only or read-and-write for each.
Inbox, calendar, spreadsheets, documents, the web.
Knowledge
What it should know before it starts: your pricing, your tone, your policies. It also remembers what it did on earlier runs.
"We never offer discounts over 15%."
Limits
The lines it can't cross without you. Limits are rules the agent is held to, set outside the instructions it reads.
No sending without approval. No more than 20 emails a day.
Handoff
Where finished work goes and who hears about problems. A good agent knows the difference between done and stuck.
Drafts to your inbox. Questions to you by message.

What agents can be.

Most useful agents fall into one of six shapes. Pick an example to load it into the composer, then change it until it's yours.

Watchers

Keep an eye on something so you don't have to, and speak up only when it changes.

Researchers

Read widely, compare what they find and come back with an answer and its sources.

Responders

Deal with what arrives: sort it, answer the routine part and pass on the rest.

Operators

Carry a task through several tools in order, the way you would with six tabs open.

Makers

Produce the same kind of document again and again, from whatever is new this time.

Teams

Several agents with one job each, passing work along and checking each other.

From a sentence to an agent you'd trust with the job.

  1. Describe the job

    Write it the way you'd brief a person. Rough is fine. If something is unclear, Agents asks you about it before going further.

  2. Read the plan

    Your sentence becomes a plan with a goal, a trigger, tools and limits, written in plain words. Change any line by editing it or by saying what's wrong.

  3. Connect what it needs

    Give the agent access to the accounts in its plan and nothing else. Each connection is read-only until you decide otherwise.

  4. Watch a test run

    The agent does the job once on your real data while you watch each step. Nothing is sent, saved or changed during a test.

  5. Switch it on

    When the test looks right, turn the agent on. It starts with the cautious setting: it asks before anything that can't be undone.

  6. Give it more room as it earns it

    Read its log, correct it when it's off and loosen its limits once you've seen it get the job right.

How much freedom an agent gets is your decision.

The same agent can work at four levels. Here is one job at each of them: a customer writes in asking for a refund on a duplicate charge.

    You can always see what it did and stop it.

    An agent you can't inspect is one you'll stop trusting, so every agent comes with the same controls.

    A log you can read

    Every run is written down step by step in plain language: what the agent saw, what it decided and why.

    Started: Monday overdue-invoice check.

    Found 4 invoices more than 14 days late.

    Skipped one. The customer emailed on Friday about a payment plan.

    Drafted 3 reminders. Waiting for your approval.

    Approvals

    Choose which actions need your yes first, such as sending, paying, deleting or publishing.

    Limits

    Cap how much it can do in a day and what it can touch. The agent can't change its own limits.

    Test runs

    Try any change against real data without sending or saving anything.

    Pause

    One switch stops an agent mid-run. It picks up where it left off when you turn it back on.

    Questions.

    Do I need to know how to code?

    No. You describe the job in plain words and edit the plan the same way. If you can brief a colleague, you can make an agent.

    How is this different from a chatbot?

    A chatbot answers when you ask. An agent has a trigger, tools and limits, so it starts by itself, does the work where the work lives and comes back to you only when it should.

    What if my idea is vague?

    Start anyway. Agents asks about the parts it can't work out, such as how often to run or who should get the result, and shows you the plan before anything runs.

    What happens when an agent gets something wrong?

    New agents ask before any action that can't be undone, so a mistake shows up as a draft you reject. The log shows where it went wrong, and you correct it in a sentence.

    What can an agent get into?

    Only the accounts you connect, at the access level you choose. You can remove a connection at any time and the agent loses that access at once.

    What happens to the idea I typed above?

    If you sign up from the plan, we save the idea with your email address so it's waiting for you as a first draft. If you don't sign up, it never leaves this page.

    You already have the idea.Make it an agent.