Your first agent

Before you begin

Every organization member can create Agents. An owner or admin assigns your spending budget; your Agent’s work stays within it.

1. Make it

Choose New agent at the top of the sidebar.

Runva creates your Agent and opens its conversation. It uses a default Computer you manage, or creates a private Computer if you haven't chosen one. Paid cloud compute starts when the Agent uses it.

There's deliberately no form here. An Agent takes its shape from what you ask it to do, and you can fill in the details once you know what you want from it.

2. Tell it what you want

Say it the way you'd say it to a person. The requests that go best name four things:

• The outcome. What should exist when it's finished?
• The sources. Which files, folders, sites, or tools does it need?
• The limits. What should it not touch, and what needs you first?
• The hand-back. What do you want to see at the end?

Something like:

Read the three PDFs I've attached, pull out every date and dollar figure, and give me one table sorted by date with a note on anything the documents disagree about. Don't change the PDFs.

Attach files to the message and Runva puts them on the Computer for the Agent to work with.

3. Watch it, or don't

The Agent's Computer is right there in the workspace, so you can see the screen, the files, and the terminal while it works. You can also close the tab. The work is on that machine, not in your browser, and it carries on without you.

You and the Agent can use the Computer at the same time. Tell it if your changes affect its work.

4. Sign it into what it needs

When the Agent needs a login, use the secure input flow to sign in yourself. Runva pauses Computer tools while the sensitive fields are present. The saved session remains on that machine.

For services with proper support, open Connections in the sidebar and add one from the catalog instead. GitHub, Linear, Notion, Stripe, Sentry, Figma, and Vercel are all there, among others.

Read Permissions and approvals before you connect anything that matters.

5. Give it a name and a job

Once you know what this Agent is for, open its settings and fill in:

• Name and Job, which is one line on what it's for.
• Instructions, which is how it should work. Anything it should always do, or never do, goes here.
• Model, Thinking, and Speed, if you want something other than the defaults. Thinking runs Low, Medium, High, Extra high, and Max, and it means the same depth whichever model you picked.
• Judge instructions, for additional restrictions on what the Agent may do. Automatic review is always enabled.

Narrow Agents work better than one that does everything. Make more when the work splits into obviously different jobs.

6. Make it repeatable

When an Agent does something well and you'll want it again, ask it to save the procedure as a Skill. When you want it to happen on its own, ask for it on a schedule and Runva makes an Automation.

Next: Threads for how conversations work, Computers for what an Agent is working on, and Skills for turning good work into something reusable.