Agents
An AI teammate that does the work and asks before anything risky
Give an agent one job, the apps it needs and your rules. It handles the routine part on its own, and when something needs your OK, it pauses and asks you in Slack, Lark or email.
What an agent is
Six things, set up once
Each agent has one job. Giving every job its own agent keeps its access and rules separate, so a mistake in one never reaches another.
Instructions
Who it is and how it does the job, in plain text you can read and change, like a brief for a new colleague.
A model
The AI that reads, thinks and writes: OpenAI, Anthropic, Gemini and others. Pick the one you trust, with your own key.
Apps
The apps it can work with, such as Stripe, GitHub, Gmail or the web, with just the tools you choose for it.
About apps →Permissions
What it may do on its own, what it must ask you first, and what it never does. Set per action.
Workflows
The jobs it does and when: every 10 minutes, every Monday at 9, or when someone asks.
About workflows →Where it reaches you
Slack, Lark, email or a webhook: where its questions and approval requests arrive.
OpenAI · gpt-6-luna
You answer our customers' emails. Look up the order before replying, keep replies short and friendly, and never promise a refund yourself.
- Look up an orderAllow
- Draft a replyAllow
- Send a replyAsk me
- Refund an orderAsk me
- Delete an emailNever
- Workflow
- Answer new emails · every 10 minutes
- Asks you in
- Slack · #support-approvals
Watch it work
One refund email, start to finish
- 1
Reads a new email
“My order #4821 arrived damaged, can I get a refund?”
- 2
Looks up order #4821
AllowShopify: delivered Tuesday, $84.00, paid by card
- 3
Drafts a reply
AllowApologises, confirms the refund, explains when it arrives
- 4
Wants to refund $84.00 and send the reply
Ask mePauses here and asks you in Slack
I'd like to refund $84.00 on order #4821 (arrived damaged) and send this reply:
Sorry your order arrived damaged! I've refunded $84.00 to your card; it should show within 5–10 days.
Your rules
Three settings for every action
Start strict, then loosen one action at a time as the agent earns your trust. Rules can depend on the details, such as allowing tracking replies but asking before refunds.
Runs on its own. For safe, routine actions such as looking up an order or answering a tracking question.
Pauses until a person approves, edits or denies it. For refunds, payments and anything sent to a customer.
Blocked outright, whatever the agent decides. For things it should never do, such as deleting customer data.
When it works
On a schedule, or when it's asked
On a schedule
Every 10 minutes, every weekday at 8, every Monday at 9. Schedules start paused until you turn them on.
When you message it
Ask it something in Chat, or mention it in Slack, and it gets to work.
From your own systems
Send it a message over the API, for example when a form is submitted or an order is placed.
Getting one
Three ways to your first agent
Start from an example
A ready-made workflow for a common job, set up in a minute: answer customer emails, approve refunds, triage incidents and more.
Browse examples →Describe a workflow
Write the job in a sentence or two. LoopEngine plans the steps, the abilities and where it asks you, then adds it to the agent you picked.
See workflows →Build it in code
Developers can write an agent in TypeScript, with their own tools and rules, and run it anywhere Node runs.
For developers →