Features

Workflows: a job in steps that runs on its own

Lay out a job as steps that hand their findings on, run it over your datasets, answer it when it asks, and run it again next time.

Say you want a profile of last quarter’s orders, then the top products and the gaps in the data side by side, then a short brief. In a chat you would ask each part in turn. A workflow is that job laid out as steps: each step is an agent’s run that takes what earlier steps found, and the whole thing runs on its own. Save it once, and next quarter run it again with a click.

Use a chat to explore. Use a workflow when you know the steps, the job repeats, or a step should run once per item of a list, such as once per region.

What you’ll need

  • One of your datasets for the workflow to start from.

Open Workflows

In the left sidebar, open Automate and choose Workflows. The editor is at the top, then My workflows and Your runs.

Make a workflow

The quickest start is one of the examples at the top: A dataset brief, One agent per group, Dig into an open question and Wait for me.

You can also have one drafted for you. In any chat, type /workflow and what it should find, such as /workflow the top five products per region and why they lead. Crunch drafts the steps from your words, the chat’s recent questions and its datasets, and opens the draft here, “Drafted from your chat”. Nothing runs until you save or run it. Any problem left, such as no dataset chosen, is listed under it.

To start from nothing, click Start from empty.

Change it in the graph

The editor shows the steps as boxes in a graph. Click a box to change it in the panel beside it. To add a step, click Input, Agent, Map, Merge or Gate after Add:; it’s linked after the selected one.

  • Input is a dataset. Under Dataset, type in “Find one of your datasets…” and choose it.
  • Agent is one agent’s run. Write its Instruction, pick what it Takes from earlier steps, and set what it Gives (findings, a table or a list). Makes turns its findings into A report, A dashboard or A sheet when it’s done.
  • Map runs one agent per item of a list, side by side. Choose the list under One run per item of, and the most items under At most (8 unless you change it, up to 10,000). Done when is all items end or any item is done.
  • Merge combines what other steps found. Union keeps one copy of each finding; Reconcile keeps conflicts with both sides.
  • Gate stops the run and asks you its Question before the steps after it go on.

As you edit, Crunch checks it: the graph says Ready to run, or how many things are left to fix, each listed below. Graph and JSON switch between the picture and its text.

Save it, and run it again

Type a name in “Name it to save it” and click Save. You’ll see “Saved · v1”. Each time you save a change to its steps, the version goes up by one; renaming doesn’t count as a change.

Under My workflows, Run again runs the saved version, Open brings it into the editor with Runs of this workflow under it, and Delete asks first (“Its runs stay.”). Run in the editor runs the saved workflow when you haven’t changed it since saving, so the run lists under Runs of this workflow. If you have unsaved changes, it asks first: Save and run saves them as a new version and runs that, and Run without saving runs what’s on screen; that run says Unsaved changes instead of the workflow’s name and isn’t listed under its runs.

Watch a run

A run opens on its own page. The header shows its state (running, waiting on you, done, failed or stopped) and the agent runs and units used; a unit measures the model work a step does. A run of a saved workflow names it and its version, with its own Run again.

Click a box to see its instruction, the Steps it took with their units, and for a map its Branches, one per item. A map where some items failed says so, such as “done · 60 of 64”, and the steps after it carry the note.

Stop ends the whole run. Findings it has made so far stay.

Answer it when it waits

A run waits for you at a gate, when a map’s list is longer than At most, or when it has used its budget. A box says “Paused at” the step and why. Click Continue (or Run the first so many, for a long list), or Stop the run. The box says what Continue does: at a budget gate it lifts that limit for the rest of the run, so the run goes on to the end and doesn’t stop for the same limit again.

What a run makes

Findings lists what each step found; a Counted: line under a figure says what it counted.

Under Use what it found, Make a report, Make a dashboard and Make a sheet turn everything found so far into one; see Pieces. What steps made themselves is under Made by this run.

To talk it over, type in “Ask about this run…” and click Ask: a new chat opens with the run’s findings ahead of your question.

Check the figures

Certify every figure, beside Run, recounts each figure independently as the run goes. It’s off unless you tick it, and adds about one more agent run per step. Certified means the figure is the value of the count it states. When a recount differs, the run pauses with Stated beside the recount, and nothing after it goes on until you choose Accept and go on or Stop the run. A figure read differently than you asked is marked Misread; one that counted something you didn’t ask for carries an Assumed: line.

On a finished step, Verify the figures counts its figures again, for about the cost of running the step once more. The marks read as in a chat: see Verify.

What if

An Agent step can split its question into parts: tick Let it split its question into parts (solve) (in beta). Once such a run ends, click a part and choose What if… to change its question; Crunch shows what runs again and what is kept with no new work, and Run the what-if starts a new run. See Work a big question in parts.

Good to know

  • Certify says a figure is the value of the computation it states, and that a different reading is said. It doesn’t promise every figure is right: read the Counted: lines for figures that matter.
  • Now and then a right figure is flagged Misread. Compare the two readings before you stop a run.

Still stuck? support@bumbleb.co

BumbleB Ask your business anything. Start free