Set up an agent, step by step

An agent is an app that does one job for your business on a schedule, for example the Finance Agent or the CRM Agent. Sasha builds it with you in one chat. This article gives the steps in order, from the first connection to the first scheduled report.

You need an admin account in Sasha. The example is the admin of Northwind Advisory, the demo business, who sets up a CRM Agent.

Before you start: give Sasha enough context

An agent is only as good as what Sasha knows about your business. Each agent lists what it needs before it is built: for example, the Content Creator Agent needs how you write, with two or three real pieces, and at least five meetings or project notes from the last month. You can see the list in the Set it up box in Agents that work for me.

When you start a set-up, Sasha first checks each item against what it really holds: your business brain, your projects and your connected tools. It then says Ready, Thin or Not ready, with the evidence for each item.

• Ready: the set-up goes on.
• Thin or Not ready: Sasha says what the agent would get wrong, and gives you one prompt for each gap. Paste each prompt into a new chat. Sasha asks you a few questions and writes the answer where every agent can use it. Then ask again to set up the agent.

You can also tell Sasha to go on without filling the gaps. Sasha then asks those questions during the set-up, and the agent starts with less to work from.

Before you start: connect the sources

An agent can read only what is connected to Sasha. Connect each source first, so that Sasha can find it during the set-up. Each agent's page says which sources it needs. For the CRM Agent, see the CRM Agent page.

There are three kinds of source.

A server for a service: Sign in or Use a key

The Add MCP Server form has tabs. Use the tab that the service tells you to use. See Connect third-party tools to Sasha for the full steps.

• Sign in. You sign in on the service's own page. Sasha keeps that sign-in for everyone. The agent can read only while the server stays signed in. If the server shows "Needs an admin to sign in again", the agent cannot read it until an admin signs in again.
• Use a key. The service gives you an address and a key. Put the key in Headers, for example Authorization=Bearer your-key. Use an equals sign, not a colon.

Choose the Server Name with care. You type it exactly in a later step, and a spelling mistake in the name is part of the name. Do not change the name after an agent uses it.

After you save a server, press Test. A working server shows its tools, for example "12 tools available." Note the names of the tools that only read. You need them later.

A GitHub synced folder for documents

A synced folder copies one folder of a GitHub repository into a Sasha project, and refreshes it every day. The agent reads the copy as part of that project. See Sync a folder from GitHub into a project.

GitHub's read-only server for code history

A synced folder copies files. It does not copy the history. To let an agent see what shipped and when, add GitHub's own read-only server with Use a key. The steps are in Connect third-party tools to Sasha, under "Read a GitHub repository's history".

Step 1: start from Agents that work for me

1. Open Agents that work for me in the side menu.
2. Click the agent with Set up next to its name, for example CRM Agent. A panel says what the agent does and what it needs.
3. Start a new chat first. Then click Set it up. Sasha puts the request, for example "Set up our CRM Agent.", in the chat box of the chat that is open.
4. Press Send.

A new chat keeps the set-up apart from your other work.

Step 2: answer Sasha's questions

Sasha reads what it already knows about your business and your connected sources. It tells you what it can see, what it cannot read yet, and what is probably missing. Then it asks only about the gaps, a few questions at a time.

Answer in your own words. When Sasha proposes a rule from your real data, check it. For example, Sasha can read Northwind's customer list and see a group that has access but does not pay. It asks what that group is before it makes a "trial" stage.

Sasha finds each source itself. It does not trust a name that you give it. If your help articles are in a different project from the one you named, Sasha says so and uses the correct project.

Step 3: check the brief

Sasha writes the agent's brief: the job, the rules, who gets the report and when. It shows you the brief before it saves it. Read it and correct what is wrong. Later, an admin can change the brief on its tab in the app, or ask Sasha in chat.

The brief has settings, a reading list and notes. The agent reads the settings and the reading list on every run. The agent does not read the notes. Notes are for the people who read the brief. A rule that the agent must follow goes in a setting or in the reading list, never only in the notes.

Step 4: choose what the agent may read

Sasha builds the agent and turns on the read tools that it can. It cannot always save these two things, so an admin does them:

• the read tools of a server that you added yourself under Add MCP Server;
• the knowledge projects that the agent may read.

Sasha tells you this, and gives you the exact lines to enter. Then:

1. Open the agent under Agents that work for me.
2. Open the ⋯ menu and choose What it may read.
3. Tick each read tool that is offered. If a tool is not offered as a tick box, type it in Tools it may use to read your figures, one per line. Write each tool as mcp__<server>__<tool>, for example mcp__northwind-customers__listCustomers. The server name must match the name in Settings exactly.
4. Enter read tools only. Never enter a tool that writes, sends or deletes.
5. Under Knowledge it may read, tick each project that the agent needs, for example the project that holds your synced help articles. A project that is not ticked is not read. Naming a project in the brief does not give access.
6. Press Save. A Confirm box opens in the Studio window. It lists what the agent may read, what it may change (usually nothing), the knowledge projects and the limits. Check them, then press Confirm.

For a picture of what an agent reads and of the Confirm dialog, see the product guide page on team roles.

Step 5: run the first report

On the Report tab, press Run now. You can also ask Sasha in the set-up chat to start the run. A Confirm box then opens in the chat. Sasha starts no run until you press Confirm.

If no read tool is saved yet, the run says "set up needed" and lists what to connect. That is correct. Go back to step 4.

Step 6: review the first report

Read the first report with the person who owns the job.

• Compare each figure with a figure that you trust.
• Look at "What it could not read". A source in that list is "not read". It is never taken as zero or as all well.
• Correct the brief where it is wrong, then press Run now again.

An admin can give the report a grade and write questions in Review, in the ⋯ menu. The next run answers each question.

Step 7: put it on a schedule

Do this only after you approve a test run. Ask Sasha in the set-up chat to put the agent on its schedule, for example "Run it every day at 06:00". The schedule then shows in Menu, Tasks, where an admin can switch it off.

If something goes wrong

For apps in general, see Use and manage apps built on your knowledge.