Getting started

An agent is a job described in your own words, plus the exact permissions it is allowed to use to do it. You write the description; the builder turns it into a definition you can read, change and run. Nothing about it is hidden from you: the steps, the tools, the limits and the cost are all on the page.

It works in a web browser and nothing else. It cannot touch files on your computer, run a terminal, use an app installed on your machine or your phone, or drive a coding session on your laptop. If a job needs one of those, the product says so while you are still describing it.

1. Say what you want done, in a sentence. No format, no fields.
2. Read the definition it writes back. It lists the steps, the tools each step may use, and the limits the runtime will enforce.
3. Change anything that is wrong by saying so. Each change is proposed, checked for whether it widens a permission, and applied against the revision you were looking at.
4. Run it. Every model call and every tool call is recorded as you watch.

Every example in this product is the same made-up company: Meridian Freight, a freight forwarding and customs clearance business in Mumbai with 11 people, 3 of them on the operations desk. It is fictional, and the gallery, the sample workspace and the sample study all describe it, so the examples agree with each other.

Its operations desk answers the same questions all week. Over the last 90 days its mailbox shows "Out for delivery alerts" asked 89 times by 1 different people, each answered with the same two lookups. That is what an agent is for: not the hard job, the repeated one.

• Shipment status questions from customers — Answers the "status consignment mfl" for you. It drafts the reply in your own words, opening "hi" and signing off "thanks", and asks you before it sends.
• Proof of delivery questions from customers — Answers the "pod mfl" for you. It drafts the reply in your own words, opening "hi" and signing off "thanks", and asks you before it sends.
• Rate enquiries to transporters — Keeps track of the "rate pune" you send out, notices who has not come back, and writes the next one in your own words, opening "hi" and signing off "thanks" for you to approve.
• Overdue invoice chasing, to customers — Keeps track of the "invoice payment overdue" you send out, notices who has not come back, and writes the next one in your own words, opening "hi" and signing off "thanks" for you to approve.

You can see this run on the sample company without signing up or connecting anything: the link is on the home page and on the first screen of setting up.

A run has two records and they are for two different readers. The run page leads with what the agent did, in sentences. Underneath, closed, is every model and tool call with exactly what went in and what came back. The first is for understanding; the second is the evidence, and it is always there.