Agents: save a job and run it again

Write down a job once, try it, save it, then run it again or put it in charge of any chat.

Some jobs you do the same way every week: a Monday sales summary, a check of last night’s orders. An agent is that job written down once. You describe it with Crunch’s help, save it, and from then on you run it with a click or put it in charge of a chat, instead of explaining it again.

What you’ll need

Optionally, the data or connections the job reads. You can add them after you make the agent.
Optionally a workspace. An agent doesn’t need one and can stand alone.

Make an agent

Open Agents under Data and connections in the left sidebar and click New agent. In a message box you can instead type /new-agent and a one-word name.
Give it a Name: letters, digits and underscores, such as weekly_sales. If you are in a workspace, you can choose it under In a workspace (optional).
Click Create. The agent starts as a draft, and its instructions open as a document with Crunch beside it. Describe the job in your own words and let Crunch write the steps with you.

To turn a chat that went well into an agent, type /save-as-agent and a name in that chat, or choose Save this chat as an agent from the / menu. You get a draft that uses what the chat used, started from what you asked there.

Try it, save it, run it

A draft is listed under Drafts on the Agents page with four buttons:

Edit reopens its instructions with Crunch beside them.
Try runs it once as written now, without saving it. Use it to see whether it does what you meant.
Save keeps it. It then moves to Your agents.
Delete throws the draft away.

Run appears only once an agent is saved. A draft can be tried but not run.

What it uses

Open a saved agent and look under Uses. It names the datasets, workspace and connections the agent reads. Click Change uses to edit them, then Save uses. Datasets are chosen with Add datasets or Change datasets; connections are the MCP servers and APIs you have added.

What it makes

Beside Run there is a Makes menu. It sets what a run produces:

Nothing: a run opens a new chat with the agent in charge, and you say what you want from it this time.
A report, a dashboard, or both: a run makes those pieces, opens the first one, and asks it with the agent’s instructions at once. See Pieces.

The choice is saved as soon as you change it. You can set it on a draft too, before you try it.

Runs and what they made

On a saved agent’s page, Runs lists the chats it ran in, and Made by its runs lists the reports and dashboards those runs produced, so you can find last week’s result.

Put an agent in charge of a chat

In any message box, click + and choose the agent under Your agents, or type /agent- and its name. From your next message the agent works the chat. See The + and / menus.

Built in agents

At the bottom of the Agents page, Built in lists agents Crunch always has, such as ones that write a report or a brief. You can’t change them. Use in a chat puts a starting request in a new message box for you to finish, and Copy to edit makes a draft of your own, named my_ plus the original name, that starts from its instructions.

How to tell it worked

A saved agent is listed under Your agents with a Run button, and after a run it shows up under Runs.

Good to know

Names take letters, digits and underscores only; spaces and dashes are turned into underscores.
Delete on a saved agent asks first, and removes every version of it. Chats that used it keep their answers.
If Crunch can’t reach your agents, the page says “Your agents can’t be reached right now.” Try again a little later.

Workspaces: keep agents together with the data they work on.
The + and / menus: attach an agent to a chat.
Pieces: the reports and dashboards an agent can make.